Cane

Open-source agentic infrastructure. Build, eval, fine-tune, and deploy AI agents.

pip install cane-ai

Started as a university lecture search system at the Frost Institute for Data Science and Computing (University of Miami), now a pip-installable platform powering cane.fyi. Also home to Softmax Gulch, a multi-agent social simulation where AI agents with persistent hierarchical memory live in a Western frontier town.

Quickstart

from cane import Cane

app = Cane()

agent = app.agents.create(
    name="Dusty",
    personality="A retired gunslinger who paints watercolors.",
    model="trinity-large-thinking",
)

agent.ingest("company_docs.pdf")

response = agent.ask("What's our return policy?")
print(response.answer)
print(response.sources)

app.serve(port=8000)

Install

# Core
pip install cane-ai

# With Google integrations (Drive, Sheets, Calendar)
pip install cane-ai[google]

# With audio transcription (Whisper)
pip install cane-ai[audio]

# Everything
pip install cane-ai[all]

CLI

cane init                  # Create .env template and data directory
cane serve                 # Start API server on :8000
cane serve --port 3000     # Custom port
cane dev                   # Dev mode with auto-reload
cane agent-create --name "Dusty" --personality "A retired gunslinger"
cane agent-list            # List all agents
cane eval Dusty --suite basic_personality

What You Get

Agents -- Create AI agents with custom personalities, knowledge bases, and tools. Webhook and MCP tool integration, sub-agent delegation, orchestrator mode, tool chaining, conversation memory, scheduled runs, per-agent API keys, embeddable widget.

RAG -- Three-stage hybrid retrieval: embedding search (ChromaDB + BGE-base-en-v1.5) -> cross-encoder reranking (ms-marco-MiniLM) -> optional LLM reranking. Multimodal ingestion (PDF, DOCX, XLSX, CSV, audio, video, images). Smart chunking with page-aware boundaries and overlap injection.

Eval -- LLM-as-a-Judge with multi-provider support (Anthropic, OpenAI, Gemini, OpenRouter). Weighted criteria, pass/warn/fail classification, latency stats, JSON schema validation, reliability scoring (A-F). Scheduled evals with webhook notifications. Personality eval with 8 dimensions and 17 test scenarios.

Fine-Tuning -- Closed-loop pipeline: eval -> mine failures -> generate training data (high-scoring results + mined corrections) -> fine-tune on OpenAI -> deploy to workspace -> re-eval to prove improvement.

Memory -- Hierarchical persistent memory for agents. Layer 1: topic cloud (~300 tokens, always in context). Layer 2: thread summaries (generated after significant conversations). Layer 3: full RAG recall triggered by agent [RECALL: query] tags.

Model Router -- Task-based routing across open source and proprietary models via OpenRouter. Trinity, Qwen (235B/30B/Coder/VL), Claude (Sonnet/Haiku). Each task type maps to the optimal model for cost and quality.

Auth -- JWT + bcrypt with role-based access (admin/owner/member). API key auth for external access. Guest/anonymous sessions.

Multi-tenant -- Single database, tenant isolation at query level. Rate limits per plan.

Stack

Package Structure

cane/
  core/          Config, database, models, migrations, security
  agents/        Memory, streaming, prompts, scheduling, conversations
  rag/           Ingestor, chunker, search, reranker, chroma, context
  eval/          Engine, judge, personality eval, failure mining, fine-tuning
  inference/     Claude client, multi-provider routing, OpenRouter
  tools/         Webhook executor, MCP client, tool registry
  auth/          JWT, roles, API keys
  api/           FastAPI app factory + all route files
  game/          Softmax Gulch models + hierarchical memory
  cli/           CLI entry point

Python API

Agent with Memory

from cane import Cane

app = Cane()
agent = app.agents.create(name="Dusty", personality="...")

conversation = agent.conversation()
conversation.send("I'm working on a project called Atlas.")
conversation.send("It's a mapping tool for AI agents.")

# Later, in a new session
conversation2 = agent.conversation()
response = conversation2.send("What was my project called?")

cloud = agent.memory.get_cloud()
print(cloud)  # topics, relationships, goals

Eval Pipeline

agent = app.agents.get("Dusty")

results = agent.eval.run(suite="basic_personality")
print(results["grade"])           # "A-"
print(results["dimension_scores"])  # {"consistency": 91, "authenticity": 83, ...}

Fine-Tuning Loop

agent = app.agents.get("Dusty")

# Mine failures, generate dataset, fine-tune, deploy
failures = agent.eval.mine_failures()
job = agent.finetune.submit(provider="openai", base_model="gpt-4o-mini")
agent.deploy(job.model_id)

new_results = agent.eval.run("basic_personality")
print(f"Before: {results['grade']} -> After: {new_results['grade']}")

Full Server

from cane import Cane

app = Cane(
    database_url="mysql+pymysql://user:pass@localhost/cane",
    chroma_path="/data/chroma",
    secret_key="your-jwt-secret",
    anthropic_api_key="sk-ant-...",
    openrouter_api_key="sk-or-...",
    enable_game=True,
)

app.serve(host="0.0.0.0", port=8000)

Environment Variables

Model Router

Task-based routing across models via OpenRouter:

Research

• First author ICLR workshop paper diagnosing retrieval failures in multimodal RAG systems
• Two ablation studies (3,000 queries across 10 conditions): cross-encoder reranking is +4-6pp, LLM adds +1-2pp, enrichment contributes nothing
• Prior work: Claude on AWS Bedrock for automated claims validation (CSAA Insurance), active learning with GroundingDINO/CLIP/YOLOv8 at 97.1% recall (RoboCanes lab)

License

Apache 2.0. See LICENSE.