What is ResearchForge?

ResearchForge is an open-source CLI that acts as a lab protocol for Claude Code and Cursor. It freezes your benchmark baseline, runs competing hypotheses in isolated git worktrees, enforces hard constraints, validates the winner through repeated runs, and ships a clean branch with a full engineering report. The coding agent proposes experiments; ResearchForge keeps evaluation deterministic.

Is ResearchForge an autonomous coding agent?

No. Claude Code and Cursor are the researchers — they read papers, form hypotheses, and write experiment patches. ResearchForge is the lab protocol: it freezes baselines, executes experiments in isolation, measures results, rejects failures, and validates winners. The AI proposes; ResearchForge evaluates.

How do I install ResearchForge?

Install with pip: pip install researchforge. Then register it with your IDE: researchforge all install --user (installs skills for Claude Code and rules for Cursor). Python 3.12+ and Git are required. No other dependencies needed.

Is ResearchForge free?

Yes. ResearchForge is free forever, Apache 2.0 licensed, with unlimited local experiments. It runs entirely on your machine — no data is sent anywhere. An optional Enterprise add-on adds shared team infrastructure, but the core CLI is fully open source.

Does ResearchForge require Claude Code or Cursor?

No. ResearchForge works in three modes: (1) Claude Code via /researchforge-start slash commands, (2) Cursor via @researchforge-start MDC rules, or (3) standalone with any AI API key — set ANTHROPIC_API_KEY, GEMINI_API_KEY, or OPENAI_API_KEY and run researchforge research synthesize.

What is the autorun command in ResearchForge?

autorun is ResearchForge's autonomous experiment loop. It runs overnight, building an experiment graph by expanding nodes with the highest potential, re-synthesizing new hypotheses from measured results each round, and stopping when it stalls, hits a target metric, or runs out of time. Command: researchforge autorun --target 0.85 --max-hours 8 --yes

How does ResearchForge isolate experiments?

Each experiment runs in a detached git worktree at the exact baseline commit, with its own Python venv or Docker container. Your working checkout is never touched. Patches are validated at import time — a patch touching a protected path is rejected before it executes.

Research outcomes

ResearchForge Case Studies

Real competitions, real benchmark improvements, and reproducible AI research workflows that turn experiments into decision-ready evidence.

🏆 $50,000 Prize Pool · Kaggle

ROGII Wellbore Geology Prediction

RMSE 15.2 → 6.1

60% error reduction on a geoscience competition

How systematically grounding every experiment in literature — and tracking every rejection with its reason — cut wellbore geology prediction error by 60% on a Kaggle competition with 6,125 teams.

60%

RMSE improvement

Experiments run

Rejected (with reasons)

~1,700 / 6,173

Rank

KaggleGeoscienceRegressionLWD logs

Read the full story →

🥉 Bronze Medal · $850k Prize Pool

ARC Prize 2026 — ARC-AGI-3

Rank #203 / ~2,226

Bronze medal zone on the hardest AI benchmark

Graph-frontier exploration (hyp-002) outperformed every hand-tuned probe approach. ResearchForge tracked the chain from base score 0.08 to 1.21 across 4 validated experiment variants — all with full lineage.

ARC-AGIFluid intelligenceLangGraphQwen3-27B

Read the full story →

Building something with ResearchForge?

If you've used ResearchForge in a competition or research project, we'd love to feature your results.

Share your results Read the blog →