#187187th of 218 in Coding agents and AI editors
arbor
Research agent that grows a hypothesis tree to optimize code against a benchmark
- Stars
- 1.1k
- License
- Apache-2.0
- Last commit
- Nov 2026
- Last release
- Jul 2026
- Language
- Python
0 on 3 Jun 20261,147 on 11 Nov 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
Arbor runs a Coordinator and Executor agent loop that proposes hypotheses, edits code in isolated git worktrees, and runs experiments against a user-supplied benchmark. Changes merge into a per-run trunk only if they clear a configurable margin on a held-out test split. It ships as a Python CLI, a keyless MCP server, and a Codex/Claude Code skill suite.
Who it is for: Engineers optimizing a repo or benchmark metric with an autonomous agent loop
Strengths
- Held-out test split gates merges, with a configurable margin against overfitting
- Each Executor works in its own git worktree; main is untouched until you merge
- Keyless mode runs inside Claude Code or Codex using the host model
- Offline replay demo and a CPU-only NumPy example task need no API key
Weaknesses
- Needs a runnable eval script, dev/test data, and a clean git repo
- Native runtime requires an LLM provider key or local model
- Headline benchmark gains are the authors' own results, not independently checked
- Keyless MCP server is an optional extra with a separate install
What it needs
- no GPU
- Needs Python >=3.10, Git, LLM provider API key or local model (Ollama)
- Models: Anthropic, OpenAI Responses API, OpenAI-compatible via LiteLLM (DeepSeek, Gemini, Qwen, vLLM, Ollama)
- port 8765
What changed
From the release notes. No flag means the note does not say so, not that the upgrade is safe.
Also in Coding agents and AI editors
See all 218| Rank | Project | Score |
|---|---|---|
| 1 | claw-codeRust CLI coding agent harness using Anthropic and OpenAI-compatible APIs | 96 out of 100 |
| 2 | t3codeDesktop, web and mobile front end for controlling local coding agent CLIs | 95 out of 100 |
| 3 | codewhaleTerminal coding agent in Rust that works with 40+ model providers | 94 out of 100 |
| #4 | ArchonYAML workflow engine that runs AI coding agents through fixed dev steps | 93 out of 100 |
| #5 | piMinimal terminal coding agent extensible with skills and packages | 92 out of 100 |