Coco Framework documentation
Build stateful, tool-calling agents in Python, with every model call and tool call passing through middleware where Coco's check attaches.
Build stateful, tool-calling agents in Python, with every model call and tool call passing through middleware where Coco's check attaches.
New to the framework? Start with the Quickstart.
How it works
create_agent compiles a state graph which calls the model, runs each tool call the model proposes through the wrap_tool_call middleware, and loops until the model answers. Every step can be checkpointed, paused with interrupt(), and resumed.
CocoGuard from the Coco SDK is one such middleware, and the framework carries no Coco enforcement code. Architecture says what the guard does with each call.
| Import | Holds |
|---|---|
coco | create_agent, init_chat_model, the middleware, and structured output |
coco_graph | StateGraph, interrupt(), Command, Send, streaming, and the checkpointers |
coco_core | Messages, tools, runnables, prompts, and CocoTracer |
coco_openai, coco_anthropic | The OpenAI and Anthropic chat models |
pip install "coco-framework[all]" installs all of them. See Install.
Get started
- Quickstart: Install the framework, run an agent with two tools, and add Coco's check.
- Install: Choose extras, see every package and its import, and set model keys.
Build agents
- Models: Switch between OpenAI and Anthropic, and configure each provider.
- Conversations: Keep a conversation in memory, in SQLite, or in PostgreSQL.
- Streaming: Stream node updates and model tokens.
- Structured output: Return a validated Pydantic object.
- Human in the loop: Pause a tool call until a person decides, and resume.
- Middleware: See the order hooks run in, and write your own.
- Built-in middleware: Add call limits, retries, fallbacks, PII redaction, and context management.
- Custom graphs: Build your own workflow with
StateGraph. - Tracing: Inspect runs in memory with
CocoTracer, with nothing sent over the network.
Understand the framework
- Architecture: The layers, the agent loop, state channels, checkpointing, tracing, and where Coco attaches.
Look something up
- Agents and middleware:
create_agent,init_chat_model, structured output, andcoco.agents.middleware. - Graphs and checkpointers:
StateGraph, control flow types, prebuilt nodes, and checkpointers. - Core: Messages, tools, runnables, prompts, and tracing.
- Model providers:
coco_openaiandcoco_anthropic. - Troubleshooting: Common errors and their fixes.