Open Agent Toolkit (OAT)

Open Agent Toolkit is an open-source toolkit for portable, provider-agnostic agent tooling and workflows.

It helps you:

• define canonical agent assets once
• sync those assets across providers
• keep nested project AGENTS.md and CLAUDE.md files aligned with pointer, symlink, or hard-copy strategies
• use provider-agnostic CLI utilities and skills
• optionally run tracked, human-in-the-loop project workflows on top

Capability Layers

flowchart TD
  BASE["Provider Sync\ncanonical assets, drift, provider views"] --> TOOLS["CLI Utilities and Skills\nbootstrap, packs, docs tooling, diagnostics"]
  TOOLS --> FLOW["Optional Workflows\ntracked projects, reviews, PR flow"]

Loading

You can adopt any layer independently.

Quick Start

pnpm install
pnpm run cli -- help
pnpm run cli -- init --scope project
pnpm run cli -- status --scope all

Useful next commands:

• pnpm run cli -- sync --scope all
• pnpm run cli -- instructions validate --strategy symlink
• pnpm run cli -- tools install
• pnpm run cli -- docs init --app-name my-docs
• pnpm run cli -- config describe
• pnpm run cli -- config dump --json
• pnpm run cli -- project status --json
• pnpm run cli -- project list --json
• pnpm run cli -- project validate-plan --project-path <path>

For local repo development:

• pnpm build
• pnpm build:docs
• pnpm lint
• pnpm type-check
• pnpm test

pnpm worktree:validate runs package builds, lint, type checks, formatting and all workspace tests, and requires a clean tree before and after. It does not build the docs site. Run pnpm build:docs separately for docs verification; CI also runs release and skill-version checks beyond the local worktree gate. The docs test task depends on upstream package builds without scheduling its own production build, whose prebuild regenerates assets used by CLI tests.

CI and the path-scoped Release Dry Run validate ready PRs targeting main, including the ready_for_review transition. Draft PRs skip those jobs. Main pushes still run CI, and Release Dry Run can still be started manually. Run the full local checks before handoff. New CI verification cancels earlier CI runs for the same ready PR. Draft events and main pushes use isolated groups; existing release workflow concurrency is unchanged.

Docs

Full documentation lives on the docs site:

• Docs Home
• Start Here
• Provider Sync
• Instruction Sync
• Agentic Workflows
• Docs Tooling
• CLI Utilities
• Reference
• Contributing

Repo Layout

• packages/cli - OAT CLI for provider sync, docs tooling, project utilities, and diagnostics
• packages/control-plane - read-only project-state library used by the CLI for structured OAT project status, listing, and recommendation data
• packages/docs-config - config helpers for OAT-powered Fumadocs apps
• packages/docs-theme - shared React components for OAT-powered Fumadocs apps
• packages/docs-transforms - remark plugins and transform bundle for OAT-powered Fumadocs apps
• apps/oat-docs - the OAT docs site
• .agents/skills - bundled OAT skills
• .oat - OAT templates, project artifacts, repo reference, and sync state

Packages

• @open-agent-toolkit/cli
• @open-agent-toolkit/control-plane
• @open-agent-toolkit/docs-config
• @open-agent-toolkit/docs-theme
• @open-agent-toolkit/docs-transforms

Release candidate verification

The Release workflow validates the exact selected tag before publication: it must resolve to the checked-out commit and be reachable from origin/main. Checks, types, tests and docs build run on the .nvmrc baseline with forced Turbo execution. Publication checks out the validated commit SHA, rather than resolving a mutable tag again, and retains its independent build and package/archive validation. PR success does not stand in for this tagged-candidate gate.

CLI test worker budget

pnpm test runs every workspace suite. Its private root command defaults VITEST_MAX_WORKERS to 2, and the qualified CLI test task alone declares that environment input. Other packages retain their existing worker selection. This bounds concurrent subprocess and filesystem work on many-core developer hosts; it preserves every test, assertion and timeout. Set the variable explicitly when measuring another CLI budget. Turbo hashes the selected value for CLI test reuse.

The budget reduces competing subprocess and filesystem work on many-core hosts. It does not promise a hosted speedup; measure the complete workload when changing it, and preserve the suite's existing timeout and failure protections.