AI Agent harnessing for durable autonomous playbooks.
The current AI agent landscape is powerful, but still fragmented and manual. We have good models, good tools, and good skills, but turning them into a reliable workflow for complex work still takes a lot of glue.
OpenPlaybooks builds the framework for autonomous playbooks — versioned, inspectable markdown artifacts that chain tasks and skills into a workflow an agent can run end to end, with checks, retries, and self-correction built into the loop.
Not a static workflow. A living playbook.
A build system for AI agents. You write playbooks as markdown files and folders; Converge compiles them into a DAG and dispatches AI agents to run it. Diverge → converge: break a big problem into independent pieces, run them in parallel, assemble the result.
- Checks, not vibes — shell-command verification (
tsc,eslint, tests). No LLM judging its own output. - Fingerprint caching — SHA-256 per node. Kill at task 47; resume from what completed.
- DAG, not context window — each TASK.md fits in one window. The runtime chains them. 670 tasks, zero lost context.
- Swap providers, not workflows — Claude, Gemini, Kimi, Qwen, Codex behind one config.
All packages are published under the @openplaybooks npm scope.
| Package | Purpose |
|---|---|
@openplaybooks/converge |
CLI for Converge — converge init, converge add, converge run |
@openplaybooks/converge-core |
Build system engine — plan, execute, verify, fix, ship |
@openplaybooks/agentfn |
Unified agent interface — switch between Claude, Kimi, Qwen, Gemini in one call |
| Package | Purpose |
|---|---|
@openplaybooks/claudefn |
Claude Code CLI as a function |
@openplaybooks/codexfn |
OpenAI Codex CLI as a function |
@openplaybooks/geminifn |
Gemini CLI as a function — with agentic loops |
@openplaybooks/kimifn |
Kimi CLI as a function — with agentic loops |
@openplaybooks/qwenfn |
Qwen CLI as a function — with agentic loops |
@openplaybooks/acpfn |
Anthropic Client Protocol (Claude Agent SDK) as a function |
@openplaybooks/openfn |
Opencode AI client as a function |
@openplaybooks/deepcodefn |
HKUDS DeepCode CLI as a function |
| Package | Purpose |
|---|---|
@openplaybooks/codets |
Fluent, indentation-aware source code emitter for TypeScript/JSX |
@openplaybooks/converge-provider-benchmark |
Deep journal analysis for comparing AI backends across playbook runs |
@openplaybooks/converge-swebench |
SWE-bench Lite runner for Converge |
@openplaybooks/converge-tbench |
Terminal-bench runner for Converge |
# 1. Install
npm install -g @openplaybooks/converge
# 2. Bootstrap a project
converge init --name=my-project --provider-template=codex
# 3. Run a built-in example, or generate one from a prompt
converge add --from-example hello-world
converge runThe five-minute walkthrough: Your first playbook.
⚠️ Converge dispatches AI agents that call LLM APIs. A playbook can consume tens of millions of tokens — pick a cheap model in the provider config.
- Source: github.com/openplaybooks-dev/converge
- Issues: openplaybooks-dev/converge/issues
- Discussions: openplaybooks-dev/converge/discussions
- npm: npmjs.com/org/openplaybooks
- Code of conduct: Contributor Covenant v2.1
- Security: report privately via GitHub Security Advisories
- Contact: minh.lucvan@ncc.asia
MIT licensed. Copyright © 2026 Converge Framework Contributors.