Start here

Coactra composes policy-aware AI agents around host-supplied models, tools, memory, workspaces, and teams.

Install and run

python3 -m pip install "coactra[agent]"

The root README contains the offline TestModel quickstart. For a real provider, pass the provider's pydantic-ai model object or model id and configure its credentials outside Coactra.

coactra doctor
coactra init my-agent
coactra validate team.json

init creates a scaffold; its generated provider-backed app still needs that provider's credentials. doctor checks the local install, and validate checks only the JSON team-spec shape.

When you need more

Need	Source
Understand ownership, scope, policy, persistence, and production limits	Architecture
Copy runnable code	Examples
Prepare a release	Release
Check exact behavior	coactra/tests/
Check exact signatures/exports	coactra/src/coactra/
Review user-visible changes	CHANGELOG.md

Optional extras are adapters, not hosted services: install agent-gateway/oauth for MCP and auth wiring, mem0 or graphiti for memory, team for SQLite directory storage, and a2a or openfga for those client adapters. Hosts provide the provider clients, credentials, and external services.

Agent runs return a RunResult; provider, tool, and usage-limit failures are terminal results with status="error". Configuration, policy, and missing-extra failures raise exceptions. Local workspace execution is disabled by default and is not a sandbox when explicitly enabled.

Code and tests are authoritative. These pages intentionally avoid copied API inventories and provider manuals.