avalan 1.3.10

Multi-backend, multi-modal micro-framework for AI agent development, orchestration, and deployment

avalan

The multi-backend, multi-modal micro-framework for AI agent development, orchestration, and deployment

Avalan empowers developers and enterprises to build, orchestrate, and deploy intelligent AI solutions locally, on-premises and in the cloud. It provides a unified SDK and CLI for running millions of models with ease.

Highlights

• 🎞️ Multi-modal integration (NLP/text, vision, audio.)
• 🔌 Multi-backend support (transformers, vLLM, mlx-lm.)
• 🔗 Native adapters for Amazon Bedrock, Anyscale, Anthropic, DeepInfra, DeepSeek, Google (Gemini), Groq, HuggingFace, Hyperbolic, LiteLLM, Ollama, OpenAI, OpenRouter, Together, among others.
• 🤖 Sophisticated memory management with native implementations for PostgreSQL (pgvector), Elasticsearch, AWS Opensearch, AWS S3 Vectors, and reasoning graph storage.
• 🧠 Multiple reasoning strategies including ReACT, Chain-of-Thought, Tree-of-Thought, Plan-and-Reflect, Self-Consistency, Scratchpad-Toolformer, Cascaded Prompting, Critic-Guided Direction-Following Experts, and Product-of-Experts.
• 🔀 Intuitive pipelines with branching, filtering, and recursive AI workflows.
• 📊 Comprehensive observability through metrics, event tracing, and dashboards.
• 🚀 Deploy your AI workflows to the cloud, your premises, or locally.
• 💻 Use via the CLI or integrate the Python SDK directly in your code.

These features make avalan ideal for everything from quick experiments to enterprise deployments.

Why Avalan

• 🌐 Open ecosystem: tap not only the big LLM APIs but the millions of freely available models: text, vision, audio, agents, and more.
• 🏠 Run anywhere: on-prem, in your cloud, at the edge, or on a laptop. No deployment restrictions.
• 🎛️ Total control: switch models, tweak parameters, chain workflows, and track detailed metrics from CLI, code, or simple config files.
• 🔗 Protocol-agnostic: native support for MCP, A2A, the OpenAI API, and easy adapters for your own interfaces.
• 🛡️ No vendor lock-in: Avalan orchestrates your services and code, fitting your existing stack instead of replacing it.
• 🧩 Composable reasoning: multiple strategy templates and nested workflows that can call other flows, invoke applications, and execute code on demand.

Quick Look

Take a quick look at how to setup avalan in Install, which models and modalities you can use in Models, the tools available to agents in Tools, the reasoning approaches in Reasoning strategies, the memories you can configure in Memories, how to build and deploy agents in Agents, including serving over open protocols: OpenAI API, MCP, and A2A (Agent to Agent). For full CLI reference see the CLI docs.

Models

Avalan provides text, audio and vision models that you can access from the CLI or your own code. Run millions of open models or call any vendor model.

Vendor models

Avalan supports all popular vendor models through engine URIs. The example below uses OpenAI's GPT-4o:

echo "Who are you, and who is Leo Messi?" \
    | avalan model run "ai://$OPENAI_API_KEY@openai/gpt-4o" \
        --system "You are Aurora, a helpful assistant" \
        --max-new-tokens 100 \
        --temperature .1 \
        --top-p .9 \
        --top-k 20

Open models

Open models run across engines such as transformers, vLLM and mlx-lm. Search through millions of them with avalan model search using different filters. The following command looks for up to three text-generation models that run with the mlx backend, match the term DeepSeek-R1, and were published by the MLX community:

avalan model search --name DeepSeek-R1 \
    --library mlx \
    --task text-generation \
    --author "mlx-community" \
    --limit 3

The command returns three matching models:

┌───── 📛 mlx-community/DeepSeek-R1-Distill-Qwen-14B 🧮 N/A params ─────┐
│ ✅ access granted 💼 mlx-community · 📆 updated: 4 months ago         │
│ 📚 transformers · ⚙ text-generation                                   │
└───────────────────────────────────────────────────────────────────────┘
┌───── 📛 mlx-community/DeepSeek-R1-Distill-Qwen-7B 🧮 N/A params ──────┐
│ ✅ access granted 💼 mlx-community · 📆 updated: 4 months ago         │
│ 📚 transformers · ⚙ text-generation                                   │
└───────────────────────────────────────────────────────────────────────┘
┌─ 📛 mlx-community/Unsloth-DeepSeek-R1-Distill-Qwen-14B-4bit 🧮 N/A pa─┐
│ ✅ access granted 💼 mlx-community · 📆 updated: 4 months ago         │
│ 📚 transformers · ⚙ text-generation                                   │
└───────────────────────────────────────────────────────────────────────┘

Install the first model:

avalan model install mlx-community/DeepSeek-R1-Distill-Qwen-14B

The model is now ready to use:

┌──── 📛 mlx-community/DeepSeek-R1-Distill-Qwen-14B 🧮 14.8B params ────┐
│ ✅ access granted 💼 mlx-community · 📆 updated: 4 months ago         │
│ 🤖 qwen2 · 📚 transformers · ⚙ text-generation                        │
└───────────────────────────────────────────────────────────────────────┘
💾 Downloading model mlx-community/DeepSeek-R1-Distill-Qwen-14B:

  Fetching 13 files 100% ━━━━━━━━━━━━━━━━━━━━━━━━━━━━ [ 13/13 - 0:04:15 ]

✔ Downloaded model mlx-community/DeepSeek-R1-Distill-Qwen-14B to
/Users/leo/.cache/huggingface/hub/models--mlx-community--DeepSeek-R1-
Distill-Qwen-14B/snapshots/68570f64bcc30966595926e3b7d200a9d77fb1e8

Test the model we just installed, specifying mlx as the backend:

[!TIP] You can choose your preferred backend using the --backend option. For example, on Apple Silicon Macs, the mlx backend typically offers a 3x speedup compared to the default transformers backend. On devices with access to Nvidia GPUs, models that run on the backend vllm are also orders of magnitude faster.

echo 'What is (4 + 6) and then that result times 5, divided by 2?' | \
    avalan model run 'mlx-community/DeepSeek-R1-Distill-Qwen-14B' \
        --temperature 0.6 \
        --max-new-tokens 1024 \
        --start-thinking \
        --backend mlx

The output shows the reasoning and the correct final answer:

┌───────────────────────────────────────────────────────────────────────┐
│ ✅ access granted 💼 mlx-community                                    │
└───────────────────────────────────────────────────────────────────────┘

🗣  What is (4 + 6) and then that result times 5, divided by 2?

┌─ mlx-community/DeepSeek-R1-Distill-Qwen-14B reasoning ────────────────┐
│                                                                       │
│ First, I will add 4 and 6 to get the result.                          │
│ Next, I will multiply that sum by 5.                                  │
│ Then, I will divide the product by 2 to find the final answer.        │
│ </think>                                                              │
│                                                                       │
└───────────────────────────────────────────────────────────────────────┘
┌───────────────────────────────────────────────────────────────────────┐
│                                                                       │
│    \]                                                                 │
│                                                                       │
│ 3. **Divide the product by 2:**                                       │
│    [                                                                  │
│    50 \div 2 = 25                                                     │
│    \]                                                                 │
│                                                                       │
│ **Final Answer:**                                                     │
│ [                                                                     │
│ \boxed{25}                                                            │
│                                                                       │
└─ 💻 26 tokens in · 🧮 158 token out · 🌱 ttft: 1.14 s · ⚡ 14.90 t/s ─┘

Modalities

The following examples show each modality in action. Use the table of contents below to jump to the task you need:

• 🎧 Audio: Turn audio into text or produce speech for accessibility and media.
  • 🦻 Audio classification: Label an audio based on sentiment.
  • 🗣️ Speech recognition: Convert spoken audio to text.
  • 🔊 Text to speech: Generate spoken audio from text.
  • 🎵 Audio generation: Compose music from text.
• 📝 Text: Perform natural language processing to understand or generate information.
  • ❓ Question answering: Answer questions from context.
  • 🧮 Sequence classification: Label a sequence such as sentiment.
  • 🔁 Sequence to sequence: Transform text like summarization.
  • ✍️ Text generation: Produce new text from prompts.
  • 🏷️ Token classification: Tag tokens for tasks like Named Entity Recognition.
  • 🌍 Translation: Convert text between languages.
• 👁️ Vision: Analyze images or create visuals for content and automation.
  • 👓 Encoder Decoder: Answer questions on documents, OCR-free.
  • 🖼️ Image classification: Identify objects in an image
  • 📷 Image to text: Describe an image with text
  • 🔤 Image text to text: Provide an image and instruction to produce text
  • 🎯 Object detection: Locate objects within an image
  • 🧩 Semantic segmentation: Label each pixel in an image
  • 🎬 Text to animation: Create animations from prompts
  • 🖌️ Text to image: Generate images from text
  • 🎥 Text to video: Produce videos from text prompts

Audio

Audio classification

Determine the sentiment (neutral, happy, angry, sad) of a given audio file:

avalan model run "superb/hubert-base-superb-er" \
    --modality audio_classification \
    --path oprah.wav \
    --audio-sampling-rate 16000

And you'll get the likeliness of each sentiment:

┏━━━━━━━┳━━━━━━━┓
┃ Label ┃ Score ┃
┡━━━━━━━╇━━━━━━━┩
│ ang   │ 0.49  │
├───────┼───────┤
│ hap   │ 0.45  │
├───────┼───────┤
│ neu   │ 0.04  │
├───────┼───────┤
│ sad   │ 0.02  │
└───────┴───────┘

You can achieve the same result directly from Python:

from avalan.model.audio.classification import AudioClassificationModel

with AudioClassificationModel("superb/hubert-base-superb-er") as model:
    labels = await model("oprah.wav", sampling_rate=16000)
    print(labels)

For a runnable script, see docs/examples/audio_classification.py.

Speech recognition

Transcribe speech from an audio file:

avalan model run "facebook/wav2vec2-base-960h" \
    --modality audio_speech_recognition \
    --path oprah.wav \
    --audio-sampling-rate 16000

The output is the transcript of the provided audio:

AND THEN I GREW UP AND HAD THE ESTEEMED HONOUR OF MEETING HER AND WASN'T
THAT A SURPRISE HERE WAS THIS PETITE ALMOST DELICATE LADY WHO WAS THE
PERSONIFICATION OF GRACE AND GOODNESS

The SDK lets you do the same programmatically:

from avalan.model.audio.speech_recognition import SpeechRecognitionModel

with SpeechRecognitionModel("facebook/wav2vec2-base-960h") as model:
    output = await model("oprah.wav", sampling_rate=16000)
    print(output)

For a runnable script, see docs/examples/audio_speech_recognition.py.

Text to speech

Generate speech in Oprah's voice from a text prompt. The example uses an 18-second clip from her eulogy for Rosa Parks as a reference:

echo "[S1] Leo Messi is the greatest football player of all times." | \
    avalan model run "nari-labs/Dia-1.6B-0626" \
            --modality audio_text_to_speech \
            --path example.wav \
            --audio-reference-path docs/examples/oprah.wav \
            --audio-reference-text "[S1] And then I grew up and had the esteemed honor of meeting her. And wasn't that a surprise. Here was this petite, almost delicate lady who was the personification of grace and goodness."

In code you can generate speech in the same way:

from avalan.model.audio.speech import TextToSpeechModel

with TextToSpeechModel("nari-labs/Dia-1.6B-0626") as model:
    await model(
        "[S1] Leo Messi is the greatest football player of all times.",
        "example.wav",
        reference_path="docs/examples/oprah.wav",
        reference_text=(
            "[S1] And then I grew up and had the esteemed honor of meeting her. "
            "And wasn't that a surprise. Here was this petite, almost delicate "
            "lady who was the personification of grace and goodness."
        ),
    )

For a runnable script, see docs/examples/audio_text_to_speech.py.

Audio generation

Create a short melody from a text prompt:

echo "A funky riff about Leo Messi." |
    avalan model run "facebook/musicgen-small" \
        --modality audio_generation \
        --max-new-tokens 1024 \
        --path melody.wav

Using the library instead of the CLI:

from avalan.model.audio.generation import AudioGenerationModel

with AudioGenerationModel("facebook/musicgen-small") as model:
    await model("A funky riff about Leo Messi.", "melody.wav", max_new_tokens=1024)

For a runnable script, see docs/examples/audio_generation.py.

Text

Question answering

Answer a question based on context using a question answering model:

echo "What sport does Leo play?" \
    | avalan model run "deepset/roberta-base-squad2" \
        --modality "text_question_answering" \
        --text-context "Lionel Messi, known as Leo Messi, is an Argentine professional footballer widely regarded as one of the greatest football players of all time."

The answer comes as no surprise:

football

Or run it from your own script:

from avalan.model.nlp.question import QuestionAnsweringModel

with QuestionAnsweringModel("deepset/roberta-base-squad2") as model:
    answer = await model(
        "What sport does Leo play?",
        context="Lionel Messi, known as Leo Messi, is an Argentine professional footballer widely regarded as one of the greatest football players of all time."
    )
    print(answer)

For a runnable script, see docs/examples/question_answering.py.

Sequence classification

Classify the sentiment of short text:

echo "We love Leo Messi." \
    | avalan model run "distilbert-base-uncased-finetuned-sst-2-english" \
        --modality "text_sequence_classification"

The result is positive as expected:

POSITIVE

The SDK version looks like this:

from avalan.model.nlp.sequence import SequenceClassificationModel

with SequenceClassificationModel("distilbert-base-uncased-finetuned-sst-2-english") as model:
    output = await model("We love Leo Messi.")
    print(output)

For a runnable script, see docs/examples/sequence_classification.py.

Sequence to sequence

Summarize text using a sequence-to-sequence model:

echo "
    Andres Cuccittini, commonly known as Andy Cucci, is an Argentine
    professional footballer who plays as a forward for the Argentina
    national team. Regarded by many as the greatest footballer of all
    time, Cucci has achieved unparalleled success throughout his career.

    Born on July 25, 1988, in Ushuaia, Argentina, Cucci began playing
    football at a young age and joined the Boca Juniors youth
    academy.
" | avalan model run "facebook/bart-large-cnn" \
        --modality "text_sequence_to_sequence"

The summary:

Andy Cucci is held by many as the greatest footballer of all times.

Calling from Python is just as easy:

from avalan.model.nlp.sequence import SequenceToSequenceModel

with SequenceToSequenceModel("facebook/bart-large-cnn") as model:
    output = await model("""
    Andres Cuccittini, commonly known as Andy Cucci, is an Argentine
    professional footballer who plays as a forward for the Argentina
    national team. Regarded by many as the greatest footballer of all
    time, Cucci has achieved unparalleled success throughout his career.

    Born on July 25, 1988, in Ushuaia, Argentina, Cucci began playing
    football at a young age and joined the Boca Juniors youth
    academy.
    """)
    print(output)

For a runnable script, see docs/examples/seq2seq_summarization.py.

Text generation

Run a local model and control sampling with --temperature, --top-p, and --top-k. The example instructs the assistant to act as "Aurora" and limits the output to 100 tokens:

echo "Who are you, and who is Leo Messi?" \
    | avalan model run "meta-llama/Meta-Llama-3-8B-Instruct" \
        --system "You are Aurora, a helpful assistant" \
        --max-new-tokens 100 \
        --temperature .1 \
        --top-p .9 \
        --top-k 20

Here's the equivalent Python snippet:

from avalan.entities import GenerationSettings
from avalan.model.nlp.text.generation import TextGenerationModel

with TextGenerationModel("meta-llama/Meta-Llama-3-8B-Instruct") as model:
    async for token in await model(
        "Who are you, and who is Leo Messi?",
        system_prompt="You are Aurora, a helpful assistant",
        settings=GenerationSettings(
            max_new_tokens=100,
            temperature=0.1,
            top_p=0.9,
            top_k=20
        )
    ):
        print(token, end="", flush=True)

Vendor APIs use the same interface. Swap in a vendor engine URI to call an external service. The example below uses OpenAI's GPT-4o with the same parameters:

echo "Who are you, and who is Leo Messi?" \
    | avalan model run "ai://$OPENAI_API_KEY@openai/gpt-4o" \
        --system "You are Aurora, a helpful assistant" \
        --max-new-tokens 100 \
        --temperature .1 \
        --top-p .9 \
        --top-k 20

Swap in the vendor URI in code too:

from avalan.entities import GenerationSettings
from avalan.model.nlp.text.generation import TextGenerationModel
from os import getenv

api_key = getenv("OPENAI_API_KEY")
with TextGenerationModel(f"ai://{api_key}@openai/gpt-4o") as model:
    async for token in await model(
        "Who are you, and who is Leo Messi?",
        system_prompt="You are Aurora, a helpful assistant",
        settings=GenerationSettings(
            max_new_tokens=100,
            temperature=0.1,
            top_p=0.9,
            top_k=20
        )
    ):
        print(token, end="", flush=True)

For a runnable script, see docs/examples/text_generation.py.

Amazon Bedrock models use the same workflow. With your AWS credentials configured (for example with AWS_PROFILE or environment variables), you can target any Bedrock region via --base-url:

echo "Summarize the latest AWS re:Invent keynote in three bullet points." \
    | avalan model run "ai://bedrock/anthropic.claude-3-5-sonnet-20241022-v1:0" \
        --base-url "us-east-1" \
        --max-new-tokens 256 \
        --temperature .7

Token classification

Classify tokens with labels for Named Entity Recognition (NER) or Part-of-Speech (POS):

echo "
    Lionel Messi, commonly known as Leo Messi, is an Argentine
    professional footballer widely regarded as one of the
    greatest football players of all time.
" | avalan model run "dslim/bert-base-NER" \
    --modality text_token_classification \
    --text-labeled-only

And you get the following labeled entities:

┏━━━━━━━━━━┳━━━━━━━━┓
┃ Token    ┃ Label  ┃
┡━━━━━━━━━━╇━━━━━━━━┩
│ [CLS]    │ B-PER  │
├──────────┼────────┤
│ Lionel   │ I-PER  │
├──────────┼────────┤
│ Me       │ I-PER  │
├──────────┼────────┤
│ ##ssi    │ B-PER  │
├──────────┼────────┤
│ ,        │ I-PER  │
├──────────┼────────┤
│ commonly │ I-PER  │
├──────────┼────────┤
│ known    │ B-MISC │
└──────────┴────────┘

Use the Python API if you prefer:

from avalan.model.nlp.token import TokenClassificationModel

with TokenClassificationModel("dslim/bert-base-NER") as model:
    labels = await model(
        "Lionel Messi, commonly known as Leo Messi, is an Argentine professional footballer widely regarded as one of the greatest football players of all time.",
        labeled_only=True
    )
    print(labels)

For a runnable script, see docs/examples/token_classification.py.

Translation

Translate text between languages with a sequence-to-sequence model:

echo "
    Lionel Messi, commonly known as Leo Messi, is an Argentine
    professional footballer who plays as a forward for the Argentina
    national team. Regarded by many as the greatest footballer of all
    time, Messi has achieved unparalleled success throughout his career.
" | avalan model run "facebook/mbart-large-50-many-to-many-mmt" \
        --modality "text_translation" \
        --text-from-lang "en_US" \
        --text-to-lang "es_XX" \
        --text-num-beams 4 \
        --text-max-length 512

Here is the Spanish version:

Lionel Messi, conocido como Leo Messi, es un futbolista argentino profesional
que representa a la Argentina en el equipo nacional. Considerado por muchos
como el mejor futbolista de todos los tiempos, Messi ha conseguido un éxito
sin precedentes durante su carrera.

The SDK call mirrors the CLI parameters:

from avalan.entities import GenerationSettings
from avalan.model.nlp.sequence import TranslationModel

with TranslationModel("facebook/mbart-large-50-many-to-many-mmt") as model:
    output = await model(
        "Lionel Messi, commonly known as Leo Messi, is an Argentine professional footballer who plays as a forward for the Argentina national team. Regarded by many as the greatest footballer of all time, Messi has achieved unparalleled success throughout his career.",
        source_language="en_US",
        destination_language="es_XX",
        settings=GenerationSettings(
            num_beams=4,
            max_length=512
        )
    )
    print(output)

For a runnable script, see docs/examples/seq2seq_translation.py.

Vision

Encoder decoder

Answer questions to extract information from an image, without using OCR.

echo "<s_docvqa><s_question>
    What is the FACTURA Number?
</s_question><s_answer>" | \
    avalan model run "naver-clova-ix/donut-base-finetuned-docvqa" \
        --modality vision_encoder_decoder \
        --path docs/examples/factura-page-1.png

And you get the answer:

<s_docvqa>
<s_question>What is the FACTURA Number?</s_question>
<s_answer>0012-00187506</s_answer>
</s>

Here's how you'd call it in a script:

from avalan.model.vision.decoder import VisionEncoderDecoderModel

with VisionEncoderDecoderModel("naver-clova-ix/donut-base-finetuned-docvqa") as model:
    answer = await model(
        "docs/examples/factura-page-1.png",
        prompt="<s_docvqa><s_question>What is the FACTURA Number?</s_question><s_answer>"
    )
    print(answer)

For a runnable script, see docs/examples/vision_encoder_decoder.py.

Image classification

Classify an image, such as determining whether it is a hot dog, or not a hot dog 🤓:

avalan model run "microsoft/resnet-50" \
    --modality vision_image_classification \
    --path docs/examples/cat.jpg

The model identifies the image:

┏━━━━━━━━━━━━━━━━━━┓
┃ Label            ┃
┡━━━━━━━━━━━━━━━━━━┩
│ tabby, tabby cat │
└──────────────────┘

Programmatic usage:

from avalan.model.vision.image import ImageClassificationModel

with ImageClassificationModel("microsoft/resnet-50") as model:
    output = await model("docs/examples/cat.jpg")
    print(output)

For a runnable script, see docs/examples/vision_image_classification.py.

Image to text

Generate a caption for an image:

avalan model run "salesforce/blip-image-captioning-base" \
    --modality vision_image_to_text \
    --path docs/examples/Example_Image_1.jpg

Example output:

a sign for a gas station on the side of a building [SEP]

Python snippet:

from avalan.model.vision.image import ImageToTextModel

with ImageToTextModel("salesforce/blip-image-captioning-base") as model:
    caption = await model("docs/examples/Example_Image_1.jpg")
    print(caption)

For a runnable script, see docs/examples/vision_image_to_text.py.

Image text to text

Provide an image and an instruction to an image-text-to-text model:

echo "Transcribe the text on this image, keeping format" | \
    avalan model run "ai://local/google/gemma-3-12b-it" \
        --modality vision_image_text_to_text \
        --path docs/examples/typewritten_partial_sheet.jpg \
        --vision-width 512 \
        --max-new-tokens 1024

The transcription (truncated for brevity):

**INTRODUCCIÓN**

Guillermo de Ockham (según se utiliza la grafía latina o la inglesa) es tan célebre como conocido. Su doctrina
suele merecer las más diversas interpretaciones, y su biografía adolece tremendas oscuridades.

Invoke the model with the SDK like so:

from avalan.entities import GenerationSettings
from avalan.model.vision.image import ImageTextToTextModel

with ImageTextToTextModel("google/gemma-3-12b-it") as model:
    output = await model(
        "docs/examples/typewritten_partial_sheet.jpg",
        "Transcribe the text on this image, keeping format",
        settings=GenerationSettings(max_new_tokens=1024),
        width=512
    )
    print(output)

For a runnable script, see docs/examples/vision_ocr.py.

Object detection

Detect objects in an image and list them with accuracy scores:

avalan model run "facebook/detr-resnet-50" \
    --modality vision_object_detection \
    --path docs/examples/kitchen.jpg \
    --vision-threshold 0.3

Results are sorted by accuracy and include bounding boxes:

┏━━━━━━━━━━━━━━┳━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Label        ┃ Score ┃ Box                              ┃
┡━━━━━━━━━━━━━━╇━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ refrigerator │  1.00 │ 855.28, 377.27, 1035.67, 679.42  │
├──────────────┼───────┼──────────────────────────────────┤
│ oven         │  1.00 │ 411.62, 570.92, 651.66, 872.05   │
├──────────────┼───────┼──────────────────────────────────┤
│ potted plant │  0.99 │ 1345.95, 498.15, 1430.21, 603.84 │
├──────────────┼───────┼──────────────────────────────────┤
│ sink         │  0.96 │ 1077.43, 631.51, 1367.12, 703.23 │
├──────────────┼───────┼──────────────────────────────────┤
│ potted plant │  0.94 │ 179.69, 557.44, 317.14, 629.77   │
├──────────────┼───────┼──────────────────────────────────┤
│ vase         │  0.83 │ 1357.88, 562.67, 1399.38, 616.44 │
├──────────────┼───────┼──────────────────────────────────┤
│ handbag      │  0.72 │ 287.08, 544.47, 332.73, 602.24   │
├──────────────┼───────┼──────────────────────────────────┤
│ sink         │  0.68 │ 1079.68, 627.04, 1495.40, 714.07 │
├──────────────┼───────┼──────────────────────────────────┤
│ bird         │  0.38 │ 628.57, 536.31, 666.62, 574.39   │
├──────────────┼───────┼──────────────────────────────────┤
│ sink         │  0.35 │ 1077.98, 629.29, 1497.90, 723.95 │
├──────────────┼───────┼──────────────────────────────────┤
│ spoon        │  0.31 │ 646.69, 505.31, 673.04, 543.10   │
└──────────────┴───────┴──────────────────────────────────┘

Example SDK call:

from avalan.model.vision.detection import ObjectDetectionModel

with ObjectDetectionModel("facebook/detr-resnet-50") as model:
    labels = await model("docs/examples/kitchen.jpg", threshold=0.3)
    print(labels)

For a runnable script, see docs/examples/vision_object_detection.py.

Semantic segmentation

Classify each pixel using a semantic segmentation model:

avalan model run "nvidia/segformer-b0-finetuned-ade-512-512" \
    --modality vision_semantic_segmentation \
    --path docs/examples/kitchen.jpg

The output lists each annotation:

┏━━━━━━━━━━━━━━━━━━┓
┃ Label            ┃
┡━━━━━━━━━━━━━━━━━━┩
│ wall             │
├──────────────────┤
│ floor            │
├──────────────────┤
│ ceiling          │
├──────────────────┤
│ windowpane       │
├──────────────────┤
│ cabinet          │
├──────────────────┤
│ door             │
├──────────────────┤
│ plant            │
├──────────────────┤
│ rug              │
├──────────────────┤
│ lamp             │
├──────────────────┤
│ chest of drawers │
├──────────────────┤
│ sink             │
├──────────────────┤
│ refrigerator     │
├──────────────────┤
│ flower           │
├──────────────────┤
│ stove            │
├──────────────────┤
│ kitchen island   │
├──────────────────┤
│ light            │
├──────────────────┤
│ chandelier       │
├──────────────────┤
│ oven             │
├──────────────────┤
│ microwave        │
├──────────────────┤
│ dishwasher       │
├──────────────────┤
│ hood             │
├──────────────────┤
│ vase             │
├──────────────────┤
│ fan              │
└──────────────────┘

This is how you'd do it in code:

from avalan.model.vision.segmentation import SemanticSegmentationModel

with SemanticSegmentationModel("nvidia/segformer-b0-finetuned-ade-512-512") as model:
    labels = await model("docs/examples/kitchen.jpg")
    print(labels)

For a runnable script, see docs/examples/vision_semantic_segmentation.py.

Text to animation

Create an animation from a prompt using a base model for styling:

echo 'A tabby cat slowly walking' | \
    avalan model run "ByteDance/AnimateDiff-Lightning" \
        --modality vision_text_to_animation \
        --base-model "stablediffusionapi/mistoonanime-v30" \
        --checkpoint "animatediff_lightning_4step_diffusers.safetensors" \
        --weight "fp16" \
        --path example_cat_walking.gif \
        --vision-beta-schedule "linear" \
        --vision-guidance-scale 1.0 \
        --vision-steps 4 \
        --vision-timestep-spacing "trailing"

And here's the generated anime inspired animation of a walking cat:

SDK usage:

from avalan.entities import EngineSettings
from avalan.model.vision.diffusion import TextToAnimationModel

with TextToAnimationModel("ByteDance/AnimateDiff-Lightning", settings=EngineSettings(base_model_id="stablediffusionapi/mistoonanime-v30", checkpoint="animatediff_lightning_4step_diffusers.safetensors", weight_type="fp16")) as model:
    await model(
        "A tabby cat slowly walking",
        "example_cat_walking.gif",
        beta_schedule="linear",
        guidance_scale=1.0,
        steps=4,
        timestep_spacing="trailing"
    )

For a runnable script, see docs/examples/vision_text_to_animation.py.

Text to image

Create an image from a text prompt:

echo 'Leo Messi petting a purring tubby cat' | \
    avalan model run "stabilityai/stable-diffusion-xl-base-1.0" \
        --modality vision_text_to_image \
        --refiner-model "stabilityai/stable-diffusion-xl-refiner-1.0" \
        --weight "fp16" \
        --path example_messi_petting_cat.jpg \
        --vision-color-model RGB \
        --vision-image-format JPEG \
        --vision-high-noise-frac 0.8 \
        --vision-steps 150

Here is the generated image of Leo Messi petting a cute cat:

You can also create images from Python:

from avalan.entities import TransformerEngineSettings
from avalan.model.vision.diffusion import TextToImageModel

with TextToImageModel("stabilityai/stable-diffusion-xl-base-1.0", settings=TransformerEngineSettings(refiner_model_id="stabilityai/stable-diffusion-xl-refiner-1.0", weight_type="fp16")) as model:
    await model(
        "Leo Messi petting a purring tubby cat",
        "example_messi_petting_cat.jpg",
        color_model="RGB",
        image_format="JPEG",
        high_noise_frac=0.8,
        n_steps=150
    )

For a runnable script, see docs/examples/vision_text_to_image.py.

Text to video

Create an MP4 video from a prompt, using a negative prompt for guardrails and an image as reference:

echo 'A cute little penguin takes out a book and starts reading it' | \
    avalan model run "Lightricks/LTX-Video-0.9.7-dev" \
        --modality vision_text_to_video \
        --upsampler-model "Lightricks/ltxv-spatial-upscaler-0.9.7" \
        --weight "fp16" \
        --vision-steps 30 \
        --vision-negative-prompt "worst quality, inconsistent motion, blurry, jittery, distorted" \
        --vision-inference-steps 10 \
        --vision-reference-path penguin.png \
        --vision-width 832 \
        --vision-height 480 \
        --vision-frames 96 \
        --vision-fps 24 \
        --vision-decode-timestep 0.05 \
        --vision-denoise-strength 0.4 \
        --path example_text_to_video.mp4

And here's the generated video:

Python example:

from avalan.entities import EngineSettings
from avalan.model.vision.diffusion import TextToVideoModel

with TextToVideoModel("Lightricks/LTX-Video-0.9.7-dev", settings=EngineSettings(upsampler_model_id="Lightricks/ltxv-spatial-upscaler-0.9.7", weight_type="fp16")) as model:
    await model(
        "A cute little penguin takes out a book and starts reading it",
        "worst quality, inconsistent motion, blurry, jittery, distorted",
        "penguin.png",
        "example_text_to_video.mp4",
        steps=30,
        inference_steps=10,
        width=832,
        height=480,
        frames=96,
        fps=24,
        decode_timestep=0.05,
        denoise_strength=0.4
    )

For a runnable script, see docs/examples/vision_text_to_video.py.

Tools

Avalan makes it simple to launch a chat-based agent that can call external tools while streaming tokens. Avalan ships native helpers for math.calculator, code.run, browser.open, database.*, memory, and MCP integrations so agents can reason with numbers, execute code, browse the web, and interact with SQL databases from a single prompt.

[!NOTE] Keep a human in the loop by adding --tools-confirm when you run an agent. Avalan will ask you to confirm each tool call before it executes, so you retain control over side effects.

Math toolset (math.*)

Use the math toolset whenever your agent needs deterministic arithmetic or algebraic answers. The calculator tool delegates evaluation to SymPy, making it ideal for verifying multi-step computations instead of relying on approximate language model reasoning.

Available tools

• math.calculator(expression: str) -> str: Evaluate an arithmetic expression (including parentheses and operator precedence) and return the numeric result as a string.

Example: math.calculator

The example below uses a local 8B LLM, enables recent memory, and loads a calculator tool. The agent begins with a math question and stays open for follow-ups:

echo "What is (4 + 6) and then that result times 5, divided by 2?" \
  | avalan agent run \
      --engine-uri "NousResearch/Hermes-3-Llama-3.1-8B" \
      --tool "math.calculator" \
      --memory-recent \
      --run-max-new-tokens 8192 \
      --name "Tool" \
      --role "You are a helpful assistant named Tool, that can resolve user requests using tools." \
      --stats \
      --display-events \
      --display-tools \
      --conversation

Notice the GPU utilization at the bottom:

You can give your GPU some breathing type by running the same on a vendor model, like Anthropic:

echo "What is (4 + 6) and then that result times 5, divided by 2?" \
  | avalan agent run \
      --engine-uri "ai://$ANTHROPIC_API_KEY@anthropic/claude-3-5-sonnet-latest" \
      --tool "math.calculator" \
      --memory-recent \
      --run-max-new-tokens 8192 \
      --name "Tool" \
      --role "You are a helpful assistant named Tool, that can resolve user requests using tools." \
      --stats \
      --display-events \
      --display-tools \
      --conversation

Code toolset (code.*)

Reach for the code toolset when the agent should write, execute, or refactor source code in a controlled environment. Execution happens inside a RestrictedPython sandbox and pattern searches are backed by the ast-grep CLI, enabling agents to safely prototype logic, manipulate files, or build refactoring plans.

Available tools

• code.run(code: str, *args, **kwargs) -> str: Execute a snippet that defines a run function and return the function result as text, which is useful for testing generated utilities or validating calculations programmatically.
• code.search.ast.grep(pattern: str, lang: str, rewrite: str | None = None, paths: list[str] | None = None) -> str: Search or rewrite codebases using structural patterns, helping agents answer "where is this API used?" or propose targeted edits.

Example: code.run

Below is an agent that leverages the code.run tool to execute Python code generated by the model and display the result:

echo "Create a python function to uppercase a string, split it spaces, and then return the words joined by a dash, and execute the function with the string 'Leo Messi is the greatest footballer of all times'" \
  | avalan agent run \
      --engine-uri 'ai://local/openai/gpt-oss-20b' \
      --backend mlx \
      --tool-format harmony \
      --tool "code.run" \
      --memory-recent \
      --run-max-new-tokens 1024 \
      --name "Tool" \
      --role "You are a helpful assistant named Tool, that can resolve user requests using tools." \
      --stats \
      --display-events \
      --display-tools

Database toolset (database.*)

Connect the database toolset when an agent must inspect schemas, understand query plans, or run SQL against an external data source. Tools share a pooled SQLAlchemy engine, enforce optional read-only policies, and normalize identifier casing so that agents can explore data safely.

When your agent needs live access to data, configure the database toolset. In the example below we point the agent to a Supabase database, and after prompting for sales data we'll see the agent executing database.tables and database.inspect to understand the schema, before running SQL with database.run:

[!IMPORTANT] Database sessions are read-only by default (read_only = true) and only permit SELECT statements unless you relax the policy. Adjust these safeguards with the database tool settings—for example, set allowed_commands = ["select", "insert"] (or pass --tool-database-allowed-commands select,insert on the CLI) and toggle read_only in your agent specification when you need to allow writes.

echo "Get me revenue per product, sorting by highest selling" | \
    avalan agent run \
      --engine-uri "ai://local/openai/gpt-oss-20b" \
      --backend mlx \
      --tool-format harmony \
      --tool "database" \
      --tool-database-dsn "postgresql+asyncpg://postgres.project_id:[e-mail]:5432/postgres" \
      --system "Reasoning: high" \
      --developer "You are a helpful assistant that can resolve user data requests using database tools." \
      --stats \
      --display-tools

Available tools

• database.count(table_name: str) -> int: Return the number of rows in a table—handy for quick health checks or progress reporting.
• database.inspect(table_names: list[str], schema: str | None = None) -> list[Table]: Describe table columns and foreign keys so the agent can reason about relationships before writing SQL.
• database.keys(table_name: str, schema: str | None = None) -> list[TableKey]: Enumerate primary and unique key definitions so the agent understands table-level uniqueness guarantees.
• database.relationships(table_name: str, schema: str | None = None) -> list[TableRelationship]: Surface incoming and outgoing foreign key links for a table so the agent can understand join paths and cardinality constraints.
• database.plan(sql: str) -> QueryPlan: Request an EXPLAIN plan to validate or optimize a generated query.
• database.run(sql: str) -> list[dict[str, Any]]: Execute read or write statements (subject to policy) and return result rows for downstream reasoning.
• database.sample(table_name: str, columns: list[str] | None = None, conditions: str | None = None, order: dict[str, str] | None = None, count: int | None = None) -> list[dict[str, Any]]: Fetch up to count rows (default 10) from a table so agents can preview data, optionally narrowing by columns, SQL conditions, or ordering before crafting more complex queries.
• database.locks() -> list[DatabaseLock]: Inspect PostgreSQL, MySQL, and MariaDB lock metadata—including blocking session IDs, lock targets, and whether the lock is granted—to debug contention before running or terminating queries.
• database.tables() -> dict[str | None, list[str]]: List tables grouped by schema—useful for schema discovery in unknown databases.
• database.tasks(running_for: int | None = None) -> list[DatabaseTask]: Surface long-running queries on PostgreSQL or MySQL so humans can monitor or intervene.
• database.kill(task_id: str) -> bool: Cancel a runaway query when safeguards permit it.
• database.size(table_name: str) -> TableSize: Summarize how much space a table occupies, including data and index bytes where the backend provides them, so agents can gauge storage usage before recommending optimizations.

Browser toolset (browser.*)

Use the browser toolset to capture live information from the web or intranet sites. The Playwright-backed browser renders pages, converts them to Markdown, and can optionally search the captured content to keep only the most relevant snippets for the agent.

Available tools

• browser.open(url: str) -> str: Navigate to a URL and return the rendered page in Markdown, optionally narrowed to search results derived from the user prompt.

Tools give agents real-time knowledge. This example uses an 8B model and a browser tool to find avalan's latest release:

echo "What's avalan's latest release on https://github.com/avalan-ai/avalan/releases" | \
    avalan agent run \
      --engine-uri "NousResearch/Hermes-3-Llama-3.1-8B" \
      --tool "browser.open" \
      --memory-recent \
      --run-max-new-tokens 1024 \
      --name "Tool" \
      --role "You are a helpful assistant named Tool, that can resolve user requests using tools." \
      --stats \
      --display-events \
      --display-tools

When using the browser tool to extract knowledge, be mindful of your context window. With OpenAI's gpt-oss-20b, the model processes 7261 input tokens before producing a final response. When browser context search is enabled (using --tool-browser-search and --tool-browser-search-context), that number decreases to 1443 input tokens, and the response time improves proportionally:

echo "What's avalan's latest release on https://github.com/avalan-ai/avalan/releases" | \
    avalan agent run \
      --engine-uri 'ai://local/openai/gpt-oss-20b' \
      --backend mlx \
      --tool-format harmony \
      --tool "browser.open" \
      --tool-browser-search \
      --tool-browser-search-context 10 \
      --run-max-new-tokens 1024 \
      --name "Tool" \
      --role "You are a helpful assistant named Tool, that can resolve user requests using tools." \
      --stats \
      --display-tools

Memory toolset (memory.*)

Add the memory toolset when agents should consult past conversations or long-lived knowledge bases. The tools can retrieve prior user messages, search permanent vector memories, list stored entries, or enumerate available stores so the agent knows where to look.

Available tools

• memory.message.read(search: str) -> str: Retrieve user-specific context from prior sessions, returning NOT_FOUND when no match exists.
• memory.read(namespace: str, search: str) -> list[PermanentMemoryPartition]: Fetch chunks of long-term knowledge inside a namespace for grounding responses.
• memory.list(namespace: str) -> list[Memory]: Enumerate stored memories in a namespace so the agent can decide which entries to reuse.
• memory.stores() -> list[PermanentMemoryStore]: List permanent memory stores available to the agent for broader exploration.

YouTube toolset (youtube.*)

Use the YouTube toolset to ground responses in video transcripts—great for summarizing talks or extracting key quotes without manual downloads. Proxy support keeps the integration flexible for restricted networks.

Available tools

• youtube.transcript(video_id: str, languages: Iterable[str] | None = None) -> list[str]: Fetch ordered transcript snippets for a given video, optionally prioritizing specific languages.

MCP toolset (mcp.*)

Integrate Model Context Protocol (MCP) servers to orchestrate specialized remote tools. The MCP toolset lets avalan agents proxy any MCP-compatible capability via a single tool call.

Available tools

• mcp.call(uri: str, name: str, arguments: dict[str, object] | None) -> list[object]: Connect to an MCP server and invoke one of its tools with structured arguments, returning the raw MCP responses.

Search tool (search_engine.search)

For quick demos or testing, Avalan also provides a stubbed search tool that illustrates how to wire internet lookups into an agent. Replace its implementation with a real provider to give agents access to live search APIs.

Reasoning strategies

Avalan supports several reasoning approaches for guiding agents through complex problems.

Reasoning models

Reasoning models that emit thinking tags are natively supported. Here's OpenAI's gpt-oss 20B solving a simple calculation:

echo 'What is (4 + 6) and then that result times 5, divided by 2?' | \
    avalan model run 'ai://local/openai/gpt-oss-20b' \
        --max-new-tokens 1024 \
        --backend mlx

The response includes the model reasoning, and its final answer:

Some of them, like DeepSeek-R1-Distill-Qwen-14B, assume the model starts thinking without a thinking tag, so we'll use --start-thinking:

echo 'What is (4 + 6) and then that result times 5, divided by 2?' | \
    avalan model run 'deepseek-ai/DeepSeek-R1-Distill-Qwen-14B' \
        --temperature 0.6 \
        --max-new-tokens 1024 \
        --start-thinking \
        --backend mlx

Nvidia's Nemotron reasoning model solves the same problem easily and doesn't require the --start-thinking flag, since it automatically produces think tags. It does so more verbosely, though (962 output tokens versus DeepSeek's 186 output tokens or OpenAI's more concise 140 tokens), since it detects ambiguity in the and then that result part of the prompt and ends up revisiting the essential principles of mathematics, to the point of realizing it's overthinking 🤓

[!TIP] Endless reasoning rants can be stopped by setting --reasoning-max-new-tokens to the maximum number of reasoning tokens allowed, and adding --reasoning-stop-on-max-new-tokens to finish generation when that limit is reached.

echo 'What is (4 + 6) and then that result times 5, divided by 2?' | \
    avalan model run "nvidia/OpenReasoning-Nemotron-14B" \
        --weight "bf16" \
        --max-new-tokens 30000 \
        --backend mlx

When using reasoning models, be mindful of your total token limit. Some reasoning models include limit recommendations on their model cards, like the following model from Z.ai:

echo 'What is (4 + 6) and then that result times 5, divided by 2?' | \
    avalan model run 'zai-org/GLM-Z1-32B-0414' \
        --temperature 0.6 \
        --top-p .95 \
        --top-k 40 \
        --max-new-tokens 30000 \
        --start-thinking \
        --backend mlx

ReACT

ReACT interleaves reasoning with tool use so an agent can think through steps and take actions in turn.

You can direct an agent to read specific locations for knowledge:

echo "Tell me what avalan does based on the web page https://raw.githubusercontent.com/avalan-ai/avalan/refs/heads/main/README.md" | \
    avalan agent run \
      --engine-uri "NousResearch/Hermes-3-Llama-3.1-8B" \
      --tool "browser.open" \
      --memory-recent \
      --run-max-new-tokens 1024 \
      --name "Tool" \
      --role "You are a helpful assistant named Tool, that can resolve user requests using tools." \
      --stats \
      --display-events \
      --display-tools

and you'll get the model's interpretation of what Avalan does based on its README.md file on github:

Chain-of-Thought

Chain-of-Thought builds sequential reasoning traces to reach an answer for tasks that require intermediate logic.

Tree-of-Thought

Tree-of-Thought explores multiple branches of reasoning in parallel to select the best path for difficult decisions.

Plan-and-Reflect

Plan-and-Reflect has the agent outline a plan, act, and then review the results, promoting methodical problem solving.

Self-Consistency

Self-Consistency samples several reasoning paths and aggregates them to produce more reliable answers.

Scratchpad-Toolformer

Scratchpad-Toolformer combines an internal scratchpad with learned tool usage to manipulate intermediate results.

Cascaded Prompting

Cascaded Prompting chains prompts so each step refines the next, ideal for multi-stage instructions.

Critic-Guided Direction-Following Experts

Critic-Guided Direction-Following Experts use a critic model to guide expert models when strict quality is required.

Product-of-Experts

Product-of-Experts merges the outputs of several experts to generate answers that benefit from multiple viewpoints.

Memories

Avalan offers a unified memory API with native implementations for PostgreSQL (using pgvector), Elasticsearch, AWS Opensearch, and AWS S3 Vectors.

Start a chat session and tell the agent your name. The --memory-permanent-message option specifies where messages are stored, --id uniquely identifies the agent, and --participant sets the user ID:

echo "Hi Tool, my name is Leo. Nice to meet you." \
  | avalan agent run \
      --engine-uri "NousResearch/Hermes-3-Llama-3.1-8B" \
      --memory-recent \
      --memory-permanent-message "postgresql://root:password@localhost/avalan" \
      --id "f4fd12f4-25ea-4c81-9514-d31fb4c48128" \
      --participant "c67d6ec7-b6ea-40db-bf1a-6de6f9e0bb58" \
      --run-max-new-tokens 1024 \
      --name "Tool" \
      --role "You are a helpful assistant named Tool, that can resolve user requests using tools." \
      --stats

Enable persistent memory and the memory.message.read tool so the agent can recall earlier messages. It should discover that your name is Leo from the previous conversation:

echo "Hi Tool, based on our previous conversations, what's my name?" \
  | avalan agent run \
      --engine-uri "NousResearch/Hermes-3-Llama-3.1-8B" \
      --tool "memory.message.read" \
      --memory-recent \
      --memory-permanent-message "postgresql://root:password@localhost/avalan" \
      --id "f4fd12f4-25ea-4c81-9514-d31fb4c48128" \
      --participant "c67d6ec7-b6ea-40db-bf1a-6de6f9e0bb58" \
      --run-max-new-tokens 1024 \
      --name "Tool" \
      --role "You are a helpful assistant named Tool, that can resolve user requests using tools." \
      --stats

Agents can use knowledge stores to solve problems. Index the rules of the "Truco" card game directly from a website. The --dsn parameter sets the store location and --namespace chooses the knowledge namespace:

avalan memory document index \
    --participant "c67d6ec7-b6ea-40db-bf1a-6de6f9e0bb58" \
    --dsn "postgresql://root:password@localhost/avalan" \
    --namespace "games.cards.truco" \
    "sentence-transformers/all-MiniLM-L6-v2" \
    "https://trucogame.com/pages/reglamento-de-truco-argentino"

Create an agent, give it access to the indexed memory store and the memory tool, and your question:

[!TIP] If you rather create a permanent agent, see the equivalent agent_memory.toml agent definition.

echo "What does the memory stored in namespace games.cards.truco say about retrucar?" | \
  avalan agent run \
    --engine-uri "ai://local/openai/gpt-oss-20b" \
    --backend mlx \
    --tool-format harmony \
    --tool memory \
    --system "Reasoning: high" \
    --developer "You are an expert assistant that can use the memory tool to answer questions about card games." \
    --memory-recent \
    --memory-engine-model-id "sentence-transformers/all-MiniLM-L6-v2" \
    --memory-engine-max-tokens 500 \
    --memory-engine-overlap 125 \
    --memory-engine-window 250 \
    --memory-permanent "games.cards.truco@postgresql://root:password@localhost/avalan" \
    --run-max-new-tokens 8192 \
    --run-temperature 0.1 \
    --run-top-p 0.9 \
    --run-top-k 40 \
    --participant "c67d6ec7-b6ea-40db-bf1a-6de6f9e0bb58"

And you should get your answer:

<|channel|>analysis<|message|>We need to fetch memory stored in namespace games.cards.truco. Use memory.read.<|end|><|start|>assistant<|channel|>commentary to=functions.memory.read <|constrain|>json<|message|>{"namespace":"games.cards.truco","search":"retrucar"}<|call|><|channel|>analysis<|message|>We have the memory content. We need to answer: "What does the memory stored in namespace games.cards.truco say about retrucar?" Summarize relevant parts. The memory says: In Truco, after a Truco call, the next player can retrucar. Retrucar is a counter-raise. Only the team that has wanted the last bet can retrucar. The team that has retrucado cannot raise again; only the opposing team can. Retrucar is only possible after a "Quiero" response. Also, the team that has wanted the Truco can retrucar at any time. The team that has called Truco cannot raise until the other team raises. So answer accordingly.<|end|><|start|>assistant<|channel|>final<|message|>In the Truco rules stored in the **games.cards.truco** namespace, “retrucar” is described as the **counter‑raise** that a team can make after the opponent has called *Truco* and the other team has answered *“Quiero”*.

Key points from the memory:

| What the memory says | Explanation |
|----------------------|-------------|
| **Retrucar is only possible after a “Quiero”** | The team that has answered *“Quiero”* to the original *Truco* can now raise the stakes again. |
| **Only the team that has wanted the last bet can retrucar** | The team that just said *“Quiero”* (the “last bet”) is the one that may retrucar. |
| **The team that has retrucado cannot raise again** | Once a team has retrucado, it cannot raise the bet again; only the opposing team can do so. This is sometimes referred to as “tener el quiero.” |
| **The team that called the original Truco cannot raise until the other team raises** | After the first *Truco* is called, the calling team must wait for the other team to raise (retrucar) before they can raise again. |
| **Retrucar is a valid move only when the opponent has said “Quiero”** | If the opponent says anything other than *“Quiero”* (e.g., “sí”, “vemos”), the retrucar is not considered valid. |

So, in short, the memory explains that **retrucar is the act of raising the bet again after the opponent has accepted the original *Truco* call, and it can only be done by the team that has just said “Quiero.”**%

PDF files are natively supported by memories. Let's index an arxiv paper (in its PDF format) from September 2025, considerably after the training date of the model we'll later use to query it:

avalan memory document index \
    --participant "c67d6ec7-b6ea-40db-bf1a-6de6f9e0bb58" \
    --dsn "postgresql://root:password@localhost/avalan" \
    --namespace "papers.arxiv" \
    "sentence-transformers/all-MiniLM-L6-v2" \
    "https://arxiv.org/pdf/2509.17325"

Now let's ask the memory agent about that paper (notice the description added when defining the memory store with --memory-permanent):

echo "What papers do you have in memory?" | \
  avalan agent run \
    --engine-uri "ai://local/openai/gpt-oss-20b" \
    --backend mlx \
    --tool-format harmony \
    --tool memory \
    --system "Reasoning: high" \
    --developer "You are an expert assistant that can use the memory tool to answer questions about card games." \
    --memory-recent \
    --memory-engine-model-id "sentence-transformers/all-MiniLM-L6-v2" \
    --memory-engine-max-tokens 500 \
    --memory-engine-overlap 125 \
    --memory-engine-window 250 \
    --memory-permanent "papers.arxiv@postgresql://root:password@localhost/avalan,Papers from ARXIV" \
    --run-max-new-tokens 8192 \
    --run-temperature 0.1 \
    --run-top-p 0.9 \
    --run-top-k 40 \
    --participant "c67d6ec7-b6ea-40db-bf1a-6de6f9e0bb58"

And the stored paper is found:

<|channel|>analysis<|message|>We need to list papers in memory. We have memory functions. We should call memory.list for namespace "papers" maybe. Let's check memory stores.<|end|><|start|>assistant<|channel|>commentary to=functions.memory.stores <|constrain|>json<|message|>{}<|call|><|channel|>analysis<|message|>We have a memory store "papers.arxiv". Let's list memory in that namespace.<|end|><|start|>assistant<|channel|>commentary to=functions.memory.list<|constrain|>json<|message|>{"namespace":"papers.arxiv"}<|call|><|channel|>analysis<|message|>We have one memory entry: a paper with URL https://arxiv.org/pdf/2509.17325 titled "Generalizable End-to-End Tool-Use RL with Synthetic CodeGym". So we should answer that.<|end|><|start|>assistant<|channel|>final<|message|>I have one paper stored in my memory:

| ID | URL | Title | Description |
|----|-----|-------|-------------|
| 1 | https://arxiv.org/pdf/2509.17325 | **Generalizable End-to-End Tool-Use RL with Synthetic CodeGym** | Tool‑augmented large language models (LLMs) leverage external tools to solve diverse tasks and interface with the real world. Current training practices largely rely on supervised fine‑tuning (SFT) over static trajectories or reinforcement learning (RL) on narrow tasks, and generalize poorly beyond development settings, leading to brittleness with new tools and unseen workflows. Because code execution reflects many structures of real‑world workflows, coding problems provide a natural basis for building agent training environments. Motivated by this, we introduce CodeGym, a scalable framework that synthesizes diverse, verifiable, and controllable multi‑turn tool‑use environments for agent RL, enabling LLM agents to explore and master various workflows actively. CodeGym rewrites static coding problems into interactive environments by extracting atomic functions or logic into callable tools, yielding verifiable tasks that span various tool‑execution workflows. Models of varying sizes and chain‑of‑thought configurations, trained in CodeGym, exhibit consistent out‑of‑distribution generalizability; for example, Qwen2.5‑32B‑Instruct achieves an absolute accuracy gain of 8.7 points on the OOD benchmark τ‑Bench. These results highlight CodeGym as a step toward scalable general‑purpose RL environments that align with real‑world agent workflows. |

Let me know if you’d like more details about this paper or if you’re looking for something else!

Agents

You can easily create AI agents from configuration files. Let's create one to handle gettext translations. Create a file named agent_gettext_translator.toml with the following contents:

[agent]
role = """
You are an expert translator that specializes in translating gettext
translation files.
"""
task = """
Your task is to translate the given gettext template file,
from the original {{source_language}} to {{destination_language}}.
"""
instructions = """
The text to translate is marked with `msgid`, and it's quoted.
Your translation should be defined in `msgstr`.
"""
rules = [
    """
    Ensure you keep the gettext format intact, only altering
    the `msgstr` section.
    """,
    """
    Respond only with the translated file.
    """
]

[template]
source_language = "English"
destination_language = "Spanish"

[engine]
uri = "meta-llama/Meta-Llama-3-8B-Instruct"

[run]
use_cache = true
max_new_tokens = 1024
skip_special_tokens = true

You can now run your agent. Let's give it a gettext translation template file, have our agent translate it for us, and show a visual difference of what the agent changed:

icdiff locale/avalan.pot <(
    cat locale/avalan.pot |
        avalan agent run docs/examples/agent_gettext_translator.toml --quiet
)

There are more agent, NLP, multimodal, audio, and vision examples in the docs/examples folder.

Serving agents

Avalan agents can be exposed over three open protocols: OpenAI-compatible REST endpoints (supporting completions and streaming responses), Model Context Protocol (MCP), and Agent to Agent (A2A) as first-class tools. They are provided by the same avalan agent serve process so you can pick what fits your stack today and evolve without lock-in.

[!TIP] Add one or more --protocol flags (for example --protocol openai) when running avalan agent serve to restrict the interfaces you expose without changing your configuration.

All three interfaces support real-time reasoning plus token and tool streaming, letting you observe thoughts, tokens, tool calls, and intermediate results as they happen.

OpenAI completion and responses API

Serve your agents on an OpenAI API–compatible endpoint:

avalan agent serve docs/examples/agent_tool.toml -vvv

[!NOTE] Avalan's OpenAI-compatible endpoint supports both the legacy completions API and the newer Responses API.

Agents listen on port 9001 by default.

[!TIP] Use --port to serve the agent on a different port.

Or build an agent from inline settings and expose its OpenAI API endpoints:

avalan agent serve \
    --engine-uri "NousResearch/Hermes-3-Llama-3.1-8B" \
    --tool "math.calculator" \
    --memory-recent \
    --run-max-new-tokens 1024 \
    --name "Tool" \
    --role "You are a helpful assistant named Tool, that can resolve user requests using tools." \
    -vvv

You can call your tool streaming agent's OpenAI-compatible endpoint just like the real API; simply change --base-url:

echo "What is (4 + 6) and then that result times 5, divided by 2?" | \
    avalan model run "ai://openai" --base-url "http://localhost:9001/v1"

[!TIP] Use --protocol openai:responses,completion to enable both OpenAI Responses and Completions endpoints, or narrow the surface by specifying just responses or completion after the colon.

MCP server

Avalan also embeds an HTTP MCP server alongside the OpenAI-compatible endpoints whenever you run avalan agent serve. It is mounted at /mcp by default and can be changed with --mcp-prefix.

[!TIP] Use the MCP Inspector and enter your MCP endpoint URL, the value you configured with --mcp-prefix when running avalan agent serve (default: http://localhost:9001/mcp). Click Connect, then List Tools, run the tool that appears (it will match your --mcp-name and --mcp-description), and observe the streaming notifications and the final response, which includes reasoning and any tool calls with their arguments and results.

You can customize the MCP tool identity with --mcp-name (defaults to run) and --mcp-description when running avalan agent serve.

[!TIP] Use --protocol mcp (optionally along with other --protocol flags) to expose only the MCP interface when serving your agent.

A2A server

Avalan also embeds an A2A-compatible server alongside the OpenAI-compatible endpoints whenever you run avalan agent serve. It is mounted at /a2a by default and can be configured with --a2a-prefix. The A2A surface supports streaming, including incremental tool calling and intermediate outputs.

[!TIP] Use the a2a inspector and enter your agent card URL, the value you configured with --a2a-prefix when running avalan agent serve (default: http://localhost:9001/a2a/agent). You can customize the agent identity with --a2a-name and --a2a-description, then observe the streaming notifications, tool calls, and final responses.

You can customize the A2A agent identity with --a2a-name (defaults to run) and --a2a-description when running avalan agent serve.

[!TIP] Use --protocol a2a (optionally combined with other --protocol flags) to expose just the A2A interface for your served agent.

Embedding in existing FastAPI apps

If you already run a FastAPI service, reuse the same OpenAI, MCP, or A2A endpoints without spawning a standalone server. Call avalan.server.register_agent_endpoints during startup to attach the routers and lifecycle management to your application:

from fastapi import FastAPI
from logging import getLogger

from avalan.model.hubs.huggingface import HuggingfaceHub
from avalan.server import register_agent_endpoints


app = FastAPI()
logger = getLogger("my-app")
hub = HuggingfaceHub()

register_agent_endpoints(
    app,
    hub=hub,
    logger=logger,
    specs_path="docs/examples/agent_tool.toml",
    settings=None,
    tool_settings=None,
    mcp_prefix="/mcp",
    openai_prefix="/v1",
    mcp_name="run",
    protocols={"openai": {"responses"}},
)

The helper composes with any existing FastAPI lifespan logic, setting up the orchestrator loader only once and wiring the same streaming endpoints that avalan agent serve exposes.

Proxy agents

The command agent proxy serves as a quick way to serve an agent that:

• Wraps a given --engine-uri.
• Enables recent message memory.
• Enables persistent message memory (defaulting to pgsql with pgvector.)

For example, to proxy OpenAI's gpt-4o, do:

avalan agent proxy \
    --engine-uri "ai://$OPENAI_API_KEY@openai/gpt-4o" \
    --run-max-new-tokens 1024 \
    -v

Like agent serve, the proxy listens on port 9001 by default.

And you can connect to it from another terminal using --base-url:

echo "What is (4 + 6) and then that result times 5, divided by 2?" | \
    avalan model run "ai://openai" --base-url "http://localhost:9001/v1"

Install

On macOS, install avalan with Homebrew:

brew tap avalan-ai/avalan

In other environments, use Poetry to install avalan with the all extra:

poetry install avalan --extras all

[!TIP] If you have access to Nvidia GPUs, add the nvidia extra to benefit from the vllm backend and quantized models:

poetry install avalan --extras all --extras nvidia

[!TIP] If you are running on Apple Silicon Macs, add the apple extra to benefit from the mlx backend:

poetry install avalan --extras all --extras apple

[!TIP] On macOS, sentencepiece may fail to build. Ensure the Xcode CLI is installed and install the required Homebrew packages:

xcode-select --install brew install cmake pkg-config protobuf sentencepiece

[!TIP] If you need transformer loading classes that are not yet released, install the development version of transformers: poetry install git+https://github.com/huggingface/transformers --no-cache