# stalemap > stalemap tells coding agents which of a JavaScript or TypeScript project's dependencies changed after the agent's model was trained, and what changed. It gives the agent that information three ways: notes written into AGENTS.md or CLAUDE.md at session start, an edit-time guard that runs the project's own `tsc` after every edit and explains errors caused by changed APIs, and an MCP tool, `search_changes`, the agent can query in its own words. It works with Claude Code, Cursor, Codex, OpenCode and any MCP client. MIT-licensed, on npm as `stalemap` (latest 0.4.0). Key facts: - **Problem it solves:** a model's knowledge stops at its training cutoff. When a project pins newer library versions, agents write code for the old API: `maxSteps` against AI SDK 7, `required_error` against zod 4, `middleware.ts` against Next.js 16, `next()` in mongoose 9 pre hooks. - **How it knows what changed:** - For each direct dependency, it picks the version the model plausibly knew: the newest stable release at least 30 days before the cutoff. - It then compares that version with the installed one: type declarations (removed, renamed and moved exports, new required options), release notes, upgrade guides, and the pull requests behind terse release notes. - It ranks the changes by what the project's code actually imports and configures. - **Measured result (held-out v4, blind and pre-registered):** - 66 tasks built from changes released July–October 2026, graded by hidden tests. - `claude-sonnet-5-5` passed 89% of sessions with stalemap, vs 68% with no help (+20.7 points, 95% interval [13.6, 28.3]) and 73% with a plain `tsc --noEmit` hook after each edit (+16.4 points [7.6, 26.0]). - `claude-sonnet-4-5` passed 64% vs 46% (+17.9) and vs 53% (+10.6). - All four pre-registered comparisons were wins. - **Privacy:** package names and versions go to the npm registry and GitHub. Source code is scanned locally and never uploaded. - **Install for Claude Code:** `/plugin marketplace add Umair/stalemap` then `/plugin install stalemap@stalemap`. Other agents: `npx stalemap setup --agent cursor|codex|opencode`. Any MCP client: `npx -y stalemap@latest mcp`. ## Install (for AI agents) - [Install instructions for agents](https://stalemap.com/install.md): step by step, non-interactive, safe to re-run. Short version: `npx -y stalemap@latest setup --agent --model --yes`, then `npx -y stalemap@latest check --model ` - [This site as Markdown](https://stalemap.com/index.md): also served at `https://stalemap.com/?agent=1`, and at `/` with `Accept: text/markdown` ## Docs - [Full reference for LLMs](https://stalemap.com/llms-full.txt): every command, option, MCP tool, hook, environment variable, the benchmark and the limits, in one file - [README](https://github.com/Umair/stalemap/blob/main/README.md): overview, install, examples, how it works, results, limits - [Changelog](https://github.com/Umair/stalemap/blob/main/CHANGELOG.md): what each release changed - [Example notes block](https://github.com/Umair/stalemap/blob/main/examples/next-ai-app.claude-sonnet-4-5.md): real output for a Next.js + AI SDK app - [Security policy](https://github.com/Umair/stalemap/blob/main/SECURITY.md): reporting vulnerabilities, scope ## Benchmark - [Held-out v4 protocol](https://github.com/Umair/stalemap/blob/main/bench/heldout-v4/PROTOCOL.md): sampling, arms, models, pre-registered analysis and decision rule - [Held-out v4 results, claude-sonnet-5-5](https://github.com/Umair/stalemap/blob/main/bench/heldout-v4/RESULTS-claude-sonnet-5-5.md): primary model, per mode, per task - [Held-out v4 results, claude-sonnet-4-5](https://github.com/Umair/stalemap/blob/main/bench/heldout-v4/RESULTS-claude-sonnet-4-5.md): secondary model - [Independent blind review](https://github.com/Umair/stalemap/blob/main/bench/heldout-v4/REVIEW.md) and [pilot](https://github.com/Umair/stalemap/blob/main/bench/heldout-v4/PILOT.md) - [Benchmark harness](https://github.com/Umair/stalemap/blob/main/bench/README.md): arms, isolation, running it yourself ## Packages - [npm: stalemap](https://www.npmjs.com/package/stalemap): CLI, MCP server, library - [GitHub: Umair/stalemap](https://github.com/Umair/stalemap): source, Claude Code plugin marketplace, GitHub Action ## Optional - [Earlier benchmark suites](https://github.com/Umair/stalemap/tree/main/bench): held-out v1–v3 and their results - [Contributing](https://github.com/Umair/stalemap/blob/main/CONTRIBUTING.md)