#267267th of 324 in Skills, plugins and prompts
autocontext
Harness that iterates agent tasks against evaluation and keeps lessons as playbooks
Documentation ↗
(opens in a new tab)Website ↗
(opens in a new tab)Repository on GitHub ↗
(opens in a new tab)
- Stars
- 1.3k
- License
- Apache-2.0
- Last commit
- Oct 2026
- Language
- Python
680 on 3 Apr 20261,316 on 30 Oct 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
autocontext runs a task from a plain-language goal through repeated generations scored by evaluation. It keeps useful lessons, discards dead ends, and writes traces, reports, playbooks and datasets to disk for later runs. It ships Python and TypeScript CLIs (autoctx), an MCP server, a Pi extension, and optional local-model distillation.
Who it is for: Engineers building self-improving agent loops with evaluation and run history
Strengths
- Filesystem-first run output (trace.jsonl, playbook.md, score.json) that can be diffed and replayed
- Works with Anthropic, OpenAI-compatible, OpenRouter, Claude CLI, Codex, Pi, and vLLM or Ollama
- Context changes are stored as candidates and served only after matched trials confirm them
- Exposes its tools to Claude Code or Cursor via autoctx serve mcp
Weaknesses
- Large surface area: many commands, scenario families and opt-in features to learn
- Python and TypeScript packages differ; some TS features lack Python parity
- No Docker image or compose file; install through uv or bun
- Reuse studies are described as infrastructure evidence only, not production promotion
What it needs
- GPU optional
- Needs uv or bun, Node.js 22.19+ (npm CLI), LLM provider or local agent such as Pi
- Models: Anthropic, OpenAI-compatible, OpenRouter, Claude CLI, Codex
Quick answers
Is autocontext open source?
Yes. Its code is on GitHub under the Apache-2.0 license.
Also in Skills, plugins and prompts
See all 324| Rank | Project | Score |
|---|---|---|
| 1 | impeccableDesign skill with 24 slash commands for AI coding agents' frontend output | 96 out of 100 |
| 2 | skillsAgent skills for planning, TDD, debugging and code review | 95 out of 100 |
| 3 | eccAgents, skills, hooks and rules bundle for Claude Code, Codex and other harnesses | 95 out of 100 |
| #4 | gstackMarkdown slash-command skills that give Claude Code planning, review, QA and ship roles | 95 out of 100 |
| #5 | cavemanCuts token use in coding agents with terse replies and compressed inputs | 95 out of 100 |