RLM Tools

Your AI coding agent spends most of its token budget just reading your code — not reasoning about it. Every grep, file read, and glob result gets dumped into the conversation. On a large codebase, that's 25-35% of your context (and cost) burned on raw data the model never needed to see.

RLM Tools gives your agent a persistent sandbox to explore code in. Data stays server-side. Only the conclusions come back.

# For Codex / generic MCP clients
uvx rlm-tools-setup --write-policy AGENTS.md

# For Claude Code
uvx rlm-tools-setup --write-policy CLAUDE.md

This auto-detects Claude/Codex on your PATH, registers the MCP server, and writes a policy block that nudges agents toward rlm-tools for repo exploration. Use CLAUDE.md for Claude Code (it reads this file automatically) or AGENTS.md for other clients.

If you prefer manual setup:

# Claude Code
claude mcp add rlm-tools -- uvx rlm-tools

# Codex
codex mcp add rlm-tools -- uvx rlm-tools

What Changes

Without RLM Tools — agent greps for import UIKit, gets 500 matches dumped into context. Reads 10 files, burns all their content as tokens. Context window fills up. Agent forgets what it was doing.

With RLM Tools — agent runs the same exploration in a server-side Python sandbox. Data stays in sandbox memory. Only the print() output enters context:

matches = grep("import UIKit")
by_module = {}
for m in matches:
    module = m["file"].split("/")[0]
    by_module.setdefault(module, []).append(m)
for module, ms in sorted(by_module.items(), key=lambda x: -len(x[1]))[:5]:
    print(f"{module}: {len(ms)} files")

500 lines of grep results become 5 lines of summary. The agent sees what it needs, nothing more.

Real-World Impact

In typical coding workflows: 25-35% context reduction. That means your agent can explore roughly 40-50% more code before hitting context limits.

In heavy exploration tasks (reading many files, broad searches), savings go much further:

Full benchmark methodology and reproduction steps: docs/benchmarks.md

How It Works

Three MCP tools. That's the entire API:

The sandbox provides built-in helpers:

• read_file(path) / read_files(paths) — Read files into variables (cached across calls)
• grep(pattern) / grep_summary(pattern) / grep_read(pattern) — Search
• glob_files(pattern) — Find files by pattern
• tree(path, max_depth) — Directory structure
• llm_query(prompt, context) — Sub-LLM analysis (optional, requires API key)

Variables persist across rlm_execute calls within a session. The agent can build up understanding incrementally — search, filter, read, analyze — without any intermediate data touching the context window.

Works With

RLM Tools is a standard MCP server. It works with any MCP-compatible client: Claude Code, Codex, Cursor, and others.

Optional Strict Mode

If your MCP client supports tool permissions or hooks, disable/block default read/grep/glob tools for repository exploration and keep rlm-tools enabled. This is the strongest way to force the token-efficient path.

{
  "mcpServers": {
    "rlm-tools": {
      "command": "uvx",
      "args": ["rlm-tools"]
    }
  }
}

# Auto-detect Claude/Codex CLIs and configure rlm-tools
uvx rlm-tools-setup

# Print JSON config snippet for other clients
uvx rlm-tools-setup --print-json

# Print policy text instead of writing it
uvx rlm-tools-setup --print-policy

uvx rlm-tools

git clone https://github.com/stefanoshea/rlm-tools.git
cd rlm-tools
uv sync
uv run rlm-tools

Configuration

Copy .env.example to .env to customize. All settings are optional — RLM Tools works out of the box with zero config.

The core exploration features (read, grep, glob, tree) require no API key. The optional llm_query() helper calls the Anthropic API for semantic analysis within the sandbox — this is the only feature that requires a key.

Security

The sandbox is read-only and restricted:

• Imports: Safe stdlib only (re, json, collections, math, etc.)
• Builtins: Blocks exec, eval, compile, __import__, breakpoint
• File access: Read-only, scoped to session directory, path traversal blocked
• Execution: Configurable per-call timeout (default 30s)
• Rate limits: Configurable max calls per session

Background

RLM Tools implements an RLM-style exploration loop: keep raw data in tool-side memory, send only compact outputs to the model. Built on the Model Context Protocol.

Development

git clone https://github.com/stefanoshea/rlm-tools.git
cd rlm-tools
uv sync --dev
pytest tests

Run comparative benchmarks (requires a local project checkout):

RLM_EVAL_PROJECT_PATH=/path/to/project pytest evals -q -s

License

MIT