File-backed autonomous task loop for Claude Code: queue one-line goals, drain them through chunked claude -p workers, with hooks, a tiered agent fleet, and memory/context MCP servers.
View on GitHub ↗claude-workflow-kit turns Claude Code into an unattended engineering loop: queue a one-line goal and it gets researched, designed, planned, executed, and reported — one clean context window per phase, every step on disk. It is four pieces that compose: a file-backed task loop, hooks that make sessions queue their own work, a tiered agent fleet, and memory/context MCP servers that give every session state awareness.
Contents: What's in the box · Why this exists · Install · Try it · The ideas that matter · Safety · License
- taskloop/ + bin/ — the queue. Tasks are markdown files with a YAML loop: block; task-add files them (with three layers of dedup), task-go files + kicks the drainer, and the drainer runs each task in fresh claude -p workers with leases, strikes, and an acceptance-command gate. - skills/autopilot/ — the conductor protocol workers follow: right-size the pipeline (trivial/small/large), one fresh subagent per phase, adversarial critics instead of human approval, checkpoint & yield. - hooks/ — prompt-time build-initiative detection, stop-time "make it runnable" queue nudges, repeated-operation crystallize nudges (hand-derived it 3×? the loop builds you a script), session-start state briefing. Installed idempotently into ~/.claude/settings.json by hooks/installhooks.py. - agents/ — a tiered subagent fleet (finder/verifier/judge/critic/dedup/tester/fixer/synth/scout/ triage-judge) plus the design guide and a template for minting your own. - mcp/ — two dependency-light MCP servers: persistent file-based memory (save/search/read/list) and context packs (memory +
From the project README.
Add the radar badge to your README — it shows your project was picked up by MCP Radar and links to this page:
[](https://mcp.liqiwa.com/s/Thrifthunter4you--claude-Autopilot-Kit.html)
AI Governance Infrastructure — open gateway demo. Self-hosted systemprompt.io gateway in a single Rust binary: policy, audit trails, and cost attribution on every AI request, with the Systemprompt Bri
jiawei686/jev-ultrafast-mcpHand a whole browser task off in one call: a decision model drives the page server-side, so a flow costs one call, not a turn per click. Ref-based element tables, code-checked assertions, zero-model m
00200200/usagetrimKeep the signal. Cut the noise. Local CLI + MCP — fold verbose tool output for Claude Code, Codex, Cursor & Desktop.
wahyunuriman999/Muse-Skill-HubOpen-source, LLM-agnostic capability runtime for AI agents — 97 skills as typed, permission-enforced MCP tools. Final certification: 23/23 gates PASS.
KernelLord/pickysteveSkill router and context picker for Claude Code, Cursor, and Aider. Auto-picks the right skill for every prompt, with prompt-injection filtering for MCP.
alibaizhanov/denselyLossless context compression for LLMs: 2x-8x fewer tokens, byte-exact reconstruction (sha256-verified). MCP server for Claude Code & Cursor.
The top new MCP servers of the week, every Monday. No spam, unsubscribe anytime.