TryAii

Embedding-based AI model router. Understands your prompt semantically and routes to the best model based on benchmarks, cost, speed, and quality.

from tryaii import DREClient, Router

router = Router()
result = router.route("Write a Python function to merge sorted arrays")

print(result.best_model)     # e.g. "anthropic/claude-opus-5.5"
print(result.best_reasoning) # "q'=0.98 (4 real of 5) | imputed: 1/5 | cost 0.2066 ($12.00/M) | ..."

client = DREClient(api_key="sk-or-...")
response = client.chat("Write a quicksort implementation")
print(response.content)

Install

pip install tryaii

The base install includes local embeddings - all-MiniLM-L6-v2 as an int8 ONNX file on onnxruntime (no PyTorch; the 23 MB model file comes in the tryaii-minilm-int8 dependency, or set TRYAII_EMBEDDING_MODEL_PATH to a copy) - no API keys needed. pip install "tryaii[centroids]" adds sentence-transformers for building fp32 centroids or running another sentence-transformers model.

Optional extras for provider integrations:

pip install tryaii[openrouter]  # Route & call models via OpenRouter (adds httpx)
pip install tryaii[openai]      # Use OpenAI embeddings instead of local (adds openai)
pip install tryaii[redis]       # Redis client for planned distributed cache (not yet implemented)
pip install tryaii[all]         # All optional integrations

Quick Start

from tryaii import Router, Priorities

router = Router()

# Route with default balanced priorities
result = router.route("Explain quantum entanglement simply")
print(result.best_model)

# Quality-first (ignore cost)
result = router.route(
    "Debug this memory leak in my Node.js app",
    priorities=Priorities(quality=5, cost=1, speed=2),
)

# Budget mode
result = router.route(
    "Summarize this email",
    priorities=Priorities.budget(),
)

Model catalog

The package ships a starter catalog: 46 well-known models from OpenAI, Anthropic, Google, DeepSeek, xAI and Mistral, scored on 15 benchmarks. It works offline with no account. Log in (free, with your tryaii.com Google account) to route on the full catalog of 319 models:

tryaii login     # device sign-in: open the printed URL in any browser and approve the code
tryaii whoami    # the signed-in account and the downloaded catalog release
tryaii logout    # remove the credentials and the downloaded catalog

After login the full catalog downloads automatically, is refreshed at most once a day, cached under ~/.tryaii/catalog/ (or TRYAII_DRE_DATA_DIR), and used only after its Ed25519 signature checks out. In code:

Router()                    # catalog="auto": full when logged in, else starter
Router(catalog="starter")   # always the packaged catalog, no network
Router(catalog="full")      # raises LoginRequiredError when not logged in

CLI

Installing the package adds a tryaii command (same surface as the Node SDK). It opens with an animated blue→red banner, then runs your command. The banner prints to stderr and auto-suppresses when output is piped, so --json stays clean.

tryaii route "Write a Python function to merge sorted arrays" --quality=5 --cost=1
tryaii eval prompts.json --output results/my-run --quality=5 --cost=1 --speed=1
tryaii models --provider anthropic        # add --json for machine-readable output
tryaii benchmarks --json
tryaii setup                               # load the embedding model + warm centroids
tryaii cachelint request.json              # pre-flight prompt-cache analysis (warn-only)
tryaii login                               # free: unlock the full model catalog

The cachelint engine also runs inside the clients: pass cache_lint="warn" to DREClient/AsyncDREClient/OpenRouterIntegration (or set TRYAII_CACHE_LINT=warn) to lint every outgoing chat request and verify cache predictions against the response usage — warn-only and fail-open.

| models | --provider <name>, --json | | benchmarks | --json | | setup / regenerate | --model <name> | | login / logout | none | | whoami | --json |

Global flags: --no-banner (or TRYAII_NO_BANNER=1), NO_COLOR=1, -v/--verbose, -V/--version. All flags work in any position and match the npm CLI.

Eval over a dataset

# Balanced run into a named folder
tryaii eval prompts.json --output results/my-run --quality=5 --cost=1 --speed=1

# Budget-aware: --max-price is the total budget for the whole dataset
tryaii eval prompts.json --output results/budget --max-price=0.10 --output-tokens=2000
tryaii eval prompts.json --output results/budget-fit --max-price=0.10 --output-tokens=2000 --budget-mode=fit-output

The input can be an array of strings or objects with prompt, optional id, and optional category. In budgeted eval, quality/cost/speed priority flags are ignored: price is the hard constraint, and the optimizer maximizes model quality within that price. --budget-mode=fit-output lowers the fixed output token estimate when the requested length cannot fit the total budget. The command writes results.jsonl, summary.json, and index.html.

OpenRouter Integration

from tryaii import Router
from tryaii.integrations import OpenRouterIntegration

router = Router()
openrouter = OpenRouterIntegration(router, api_key="sk-or-...")

response = openrouter.chat("Write a quicksort implementation")
print(response.model_used)  # Auto-selected best model
print(response.content)     # Actual response

OpenAI Embeddings

from tryaii import Router
from tryaii.embeddings import OpenAIEmbeddingProvider

router = Router(
    embedding_provider=OpenAIEmbeddingProvider(),
)

result = router.route("Summarize this architecture decision")
print(result.best_model)

Install the OpenAI client first:

pip install tryaii[openai]

Data and privacy

Without an app key, nothing leaves your machine: the monitor writes its records under ~/.tryaii/monitor/ and that is all. With TRYAII_API_KEY set, it sends metadata about each LLM call to your tryaii.com account: model names, token counts, estimated costs, timings, the file (relative to your project) and function that made the call, and a task profile (how similar the prompt is to public benchmark categories). It never sends prompt or response text, embeddings, provider API keys or file contents. The prompt hash it sends is keyed with a secret that stays on your machine. Metadata sent to your own Langfuse (TRYAII_LANGFUSE=on) goes from your app straight to your Langfuse, never through TryAii; while the export is on, uploads to TryAii carry an integrations field naming the integrations in use.

• See what is sent and switch it off: tryaii monitor upload / tryaii monitor upload off, or set TRYAII_TELEMETRY=0 or DO_NOT_TRACK=1.
• Call records are deleted after 90 days; a deleted app is restorable for 30 days, then removed for good.
• Full text: privacy policy, "Command-line tool and SDK" and PRIVACY.md. Questions or deletion requests: support@tryaii.com.

License

Apache 2.0