Summary

FastContext is a lightweight, delegated repository explorer for coding agents, designed to optimize token usage and reasoning accuracy. It separates broad exploration tasks from main agents, improving the score-token tradeoff and providing focused evidence for AI coding assistants.

How It Works

Delegates natural-language context queries to a dedicated subagent. Utilizes read-only tools (Read, Glob, Grep) with parallel execution for efficient exploration. Returns compact file-line citations as focused evidence, preventing main agent context pollution. Trains dedicated exploration models (4B-30B) via SFT and RL.

Quick Start & Requirements

Requires Python 3.12+ and uv package management. Install CLI: uv tool install .. Development: uv sync --all-groups. Needs an OpenAI-compatible chat endpoint (configure BASE_URL, MODEL, API_KEY). CLI example: fastcontext --query "...". Programmatic use via make_fastcontext_agent. Links: arXiv paper [📄 arXiv], model weights [🤗 Model].

Highlighted Details

• Up to +5.5 score improvement on SWE-bench variants.
• Reduces main agent token use by up to 60.3%.
• Trained models (FC-4B-RL) show competitive performance.
• Accurately recovers patch-relevant files/symbols.

Maintenance & Community

Paper and model weights released June 15, 2026. No explicit community channels or detailed maintenance status provided.

Licensing & Compatibility

The README does not specify the software license, requiring further investigation for commercial use or integration into closed-source projects.

Limitations & Caveats

Strictly for repository exploration, not code modification. Tool outputs are capped for responsiveness. Precise exploration queries are recommended.