Charter 0.2.0

Any model LangChain can build, not only Anthropic.

llm:
  provider: openai              # or google_genai, bedrock, ollama, groq, …
  api_key: ${OPENAI_API_KEY}
  base_url: ${LOCAL_LLM_URL}    # optional: an OpenAI-compatible server you host

llm.provider was Literal["anthropic"], so every other provider was refused at parse time. Models are built with init_chat_model now, which means the providers Charter serves are the ones LangChain serves, and a new one there needs no change here.

• api_key is optional. A local runtime authenticates nothing, and bedrock and vertex read credentials from elsewhere.
• base_url reaches vLLM, LM Studio, llama.cpp, or a provider on a non-default host.
• Extras for the common integrations: boundflow-charter[openai], [google], [bedrock], [ollama], [groq].

The model must report token usage. BoundFlow refuses to run one that doesn't, since a cost cap it cannot price is not a cap.

Also

The README is rebuilt around a quickstart you can paste: two YAML files, four commands, no cloning. deploy/local.compose.yml brings up a control plane locally, so nothing waits on an invite. The e2e suite deletes the tenants it creates, and a stuck test now fails instead of hanging.

Still pre-alpha. The configuration format will change.

Charter 0.1.0

The first release. Charter is a declarative layer over deepagents and the BoundFlow control plane: an agent is a versioned YAML file, its budgets and lifecycle rules are policy, and the control plane holds a human's decision for as long as it takes.

pip install boundflow-charter

What's here

• Agents as configuration — objective, MCP tools, approval gates and the answer shape in v1.yaml, immutable and versioned. Budgets, authority and lifecycle rules live beside it as policy you can change without cutting a version.
• Durable human gates — a run parks at an approval or a question and resumes on another worker, days later, with what it had discovered, spent and been told. charter approve / reject / answer, each recorded with an actor and a reason.
• Lifecycle policy — pause, cool down, or roll back to a known-good version off the agent's own cost, failure and rejection metrics.
• charter ui — a local console for the fleet, its open decisions and its holds.
• Traces — every model and tool call to an OTLP backend of your own; prompts never reach the control plane.
• A control plane in two commands — deploy/local.compose.yml runs one locally, or use BoundFlow Cloud.

Status

Pre-alpha. The design is settled enough to argue with; the configuration format will change. Feedback and bug reports welcome.

Needs boundflow>=0.6.0 and Python 3.10+. pip install --pre boundflow-charter tracks main, published on every green build.