FastContext Integrations

Your coding agent is wasting tokens. In GPT-5.4 trajectories, reading and searching account for 56% of all tool-use turns and 47% of the main agent's total tokens — just to locate the relevant code. FastContext offloads that entirely to a dedicated subagent, so your main agent receives clean file:line citations instead of a long trail of exploratory reads.

The result: up to +5.5% accuracy and up to 60% fewer tokens on SWE-bench benchmarks.

This repo is the MCP glue that wires FastContext into every major editor with one click.

Image created using Nano Banana

fastcontext_explore("where are webhook signatures verified?")
→  src/auth/webhook.py:42-61
→  config/secrets.py:18

Your agent reads those two ranges. Done.

The model

FastContext-1.0 is a model family purpose-trained for repository exploration by Microsoft Research (arXiv:2606.14066). It is not a general LLM asked to search code — it is trained end-to-end on exploration trajectories using SFT then refined with task-grounded RL (GRPO), with rewards based on file- and line-level F1.

At each turn it issues parallel READ / GLOB / GREP tool calls, refines based on observations, and stops with a compact <final_answer> citation block. Nothing more enters the main agent's context.

Model family

All variants support up to 262K token context.

The compact 4B-RL explorer can outperform the larger 30B-SFT — e.g. on SWE-bench Pro with GLM-5.1 it reaches 22.5 vs. 20.0 while using fewer tokens.

Where to download

• LM Studio — search FastContext in the model browser. Pick FC-4B-SFT or FC-4B-RL for consumer hardware; use MLX builds on Apple Silicon.

• HuggingFace — microsoft/FastContext-1.0-4B-SFT, microsoft/FastContext-1.0-4B-RL, microsoft/FastContext-1.0-30B-SFT.

• Ollama / llama.cpp — any GGUF community conversion; search HuggingFace for FastContext GGUF.

Once loaded, copy the model ID exactly as shown by your runtime and paste it into --model.

Why it's fast

• Small by design: a 4B model laser-focused on one task beats a 70B generalist at it.

• Parallel tool calls in a single turn: covers multiple search hypotheses at once.

• Local and private: no code leaves your machine, no API cost per search.

Related MCP server: quick-search

Install

Claude Code (no button — one command):

claude mcp add fastcontext -- uvx --from git+https://github.com/LIVELUCKY/fastcontext-integrations fastcontext-mcp \
  --base-url http://localhost:1234/v1 --model your-model-id --api-key lm-studio

After clicking a button or running the command, set --model to the exact ID your runtime shows for the loaded model. Using a remote API? Keep the key secure — see docs/SETUP.md#secure-api-keys.

Prerequisites (once)

# 1. uv (the Python tool runner)
curl -LsSf https://astral.sh/uv/install.sh | sh

# 2. the FastContext explorer CLI on your PATH
uv tool install git+https://github.com/microsoft/fastcontext

# 3. a FastContext model loaded in an OpenAI-compatible runtime
#    e.g. LM Studio: search "FastContext", download FC-4B-SFT or FC-4B-RL,
#    Developer tab → Start Server (serves http://localhost:1234/v1, no API key needed)

No clone, no absolute paths, no environment variables: the server runs via uvx straight from this repo and takes its connection from the --base-url / --model / --api-key args. Full details in docs/SETUP.md.

Per-editor setup

Click the Install in VS Code button above — it registers the server directly in VS Code, which Copilot agent mode uses. Or copy examples/vscode.mcp.json into your project's .vscode/mcp.json (top-level key is servers, not mcpServers). Enable agent mode — fastcontext_explore appears in the tool picker. Add the usage guidance to .github/copilot-instructions.md.

Run the claude mcp add command above, or copy examples/claude-code.mcp.json to your project root as .mcp.json. Append prompts/fastcontext-usage.md to your CLAUDE.md.

Prefer the upstream-style skill (the CLI directly, reads env vars instead of args)? See examples/claude-code-skill/SKILL.md.

Add examples/codex.config.toml to ~/.codex/config.toml (header is [mcp_servers.fastcontext] — underscore). Append the usage guidance to your AGENTS.md.

Click Add to Cursor above, or copy examples/cursor.mcp.json into .cursor/mcp.json. Add the usage guidance as a .cursor/rules/fastcontext.mdc rule.

Merge examples/cline.mcp.json into Cline's MCP settings (autoApprove is pre-set for the read-only tool).

Copy examples/windsurf.mcp.json to ~/.codeium/windsurf/mcp_config.json (global) or merge into your project's .windsurf/mcp.json (local). The format is the same mcpServers object used by Cursor and Claude Code. Add the usage guidance as a Windsurf rule.

Any MCP client: register the command uvx --from git+https://github.com/LIVELUCKY/fastcontext-integrations fastcontext-mcp --base-url ... --model ....

Any shell-capable agent without MCP: install the FastContext CLI and run fastcontext -q "<question>" --citation directly (reads BASE_URL/MODEL/API_KEY from the environment). Guidance: prompts/fastcontext-usage.md.

Make the agent actually delegate

Add prompts/fastcontext-usage.md to your agent's instructions. Without it, agents tend to ignore the tool or re-scan the repo after calling it — which erases the savings. (Where it goes per client.)

Updating

uvx caches the server by commit and does not auto-update. When a new version lands, the server logs update available: vX.Y.Z on startup (visible in your client's MCP logs). Pull it with one command:

uvx --refresh --from git+https://github.com/LIVELUCKY/fastcontext-integrations fastcontext-mcp --help

Then restart your client. (Or uv cache clean fastcontext-mcp to force a rebuild on next launch.)

Verify

./scripts/fastcontext-check.sh /path/to/any/repo \
  --base-url http://localhost:1234/v1 --model your-model-id

What's in here

fastcontext_mcp.py        zero-dependency MCP server (connection via args)
pyproject.toml            makes it runnable as `uvx --from git+<repo> fastcontext-mcp`
examples/                 copy-paste config per editor (+ optional Claude skill)
prompts/                  the "when/how to delegate" usage prompt
scripts/                  make-install-buttons.py (regenerate badges), fastcontext-check.sh
docs/                     SETUP.md, TROUBLESHOOTING.md

Credits & license

FastContext is by Microsoft Research, MIT-licensed (github.com/microsoft/fastcontext, arXiv:2606.14066). The optional Claude skill and the usage prompt are adapted from that repo. This integration layer is MIT-licensed (see LICENSE). Not affiliated with or endorsed by Microsoft.