driftwatch

driftwatch

Dependency migration intelligence for AI coding agents.

Answers one question, precisely and with citations:

*This library moved from version A to version B. What broke, and what edits

does my code need?*

New here? Start with the Beginner's Guide.

This repo also holds the source of CMM Research's data Actors on Apify: job, hospital, funding and Shopify data APIs.

Install (30 seconds)

Add it to Claude Desktop, Claude Code, Cursor, or any MCP client:

{

"mcpServers": { "driftwatch": { "command": "npx", "args": ["-y", "driftwatch-mcp"] } } }

That is the whole setup. No API key, no account, no payment, no config. The engine runs locally on your machine and reads only public data — npm, PyPI, GitHub Releases, and OSV.dev.

Two tools appear in your agent:

Optionally set ANTHROPIC_API_KEY to add LLM-synthesized migration steps on top of the deterministic results. It works fully without one.

Why this exists

Every LLM is frozen at a training cutoff. Libraries are not. When an agent writes code against a version newer than its cutoff, it confidently emits an API that no longer exists — and then burns roughly six failed build-fix iterations converging on the truth.

The SDKProof benchmark (July 2026) measured this: models score 80/100 on Prisma v7 (the v6 PrismaClient pattern was removed) and 90/100 on Vercel AI SDK v5+ (parameters renamed, maxSteps deleted).

Without driftwatch:  ~6 failed iterations x ~15k tokens  ->  $0.30-$1.50 + 10-20 min

With driftwatch: 1 call -> $0.05

A 6–30x return the buyer computes for themselves. No trust required.

Quick start

npm install

cp .env.example .env npm start

# The product

curl "localhost:4021/v1/delta?ecosystem=npm&name=react&from=18.2.0&to=19.0.0"

Free safety check -- catches hallucinated and typosquatted packages

curl "localhost:4021/v1/check?ecosystem=npm&name=recat"

Runs with payments off and paid AI features off. Costs nothing.

What you get

{

"package": "react", "from": "18.2.0", "to": "19.0.0", "jump": { "kind": "major", "majorsCrossed": 1, "releasesInRange": 583 }, "tier": "evidence", "breakingChanges": [ { "summary": "Removed: ReactDOM.render, ReactDOM.hydrate ...", "version": "19.0.0", "confidence": "medium", "symbols": ["ReactDOM", "render", "hydrate"], "citations": [{ "kind": "release-note", "url": "https://github.com/..." }] } ], "advisories": [], "citations": [ /* every source we relied on */ ] }

Every claim links to a primary source. We publish facts and short citations — never wholesale documentation.

Architecture

                 ┌──────────────────────────────────────────┐

│ FREE PUBLIC SOURCES (no licensed data) │ │ npm · PyPI · GitHub Releases · OSV.dev │ └────────────────────┬─────────────────────┘ │ ┌───────────▼───────────┐ │ ENGINE │ │ ┌──────────────────┐ │ │ │ deterministic │ │ always on, $0 │ │ extraction │ │ │ └────────┬─────────┘ │ │ ┌────────▼─────────┐ │ │ │ LLM synthesis │ │ OPTIONAL, capped │ │ (off by default) │ │ │ └────────┬─────────┘ │ └───────────┼────────────┘ │ ┌───────────▼───────────┐ │ SQLite PERMANENT CACHE│ ← the margin │ + revenue ledger │ └───────────┬───────────┘ │ ┌───────────────────────┼───────────────────────┐ │ │ │ ┌───────▼───────┐ ┌────────▼────────┐ ┌────────▼────────┐ │ MCP server │ │ HTTP API │ │ x402 layer │ │ (stdio) │ │ + OpenAPI │ │ (Base, USDC) │ │ │ │ │ │ │ │ DISTRIBUTION │ │ REVENUE │ │ OPTIONALITY │ └───────────────┘ └─────────────────┘ └─────────────────┘

The strategy in one line: MCP has the users, the API has the revenue, x402 is cheap positioning. See DECISION.md for why, and MARKET_RESEARCH.md for the measured evidence.

Endpoints

Ecosystems: npm, PyPI.

MCP server

The distribution channel. Two tools: get_migration_delta and check_package.

Published as driftwatch-mcp — see Install above for the one-block setup.

To run it from a clone instead of npm:

{

"mcpServers": { "driftwatch": { "command": "node", "args": ["--experimental-strip-types", "/path/to/driftwatch/src/mcp/server.ts"] } } }

Runs the engine locally by default — no network calls to us, no payment. Set DRIFTWATCH_REMOTE_URL to point it at a hosted instance instead.

No native dependencies required. better-sqlite3 is optional; if it cannot build on your machine the server falls back to a plain JSON cache and works identically. A failed native build is the most common reason MCP servers die on install, and this one survives it.

Commands

npm start           # run the API server

npm run dev # run with auto-reload npm run mcp # run the MCP server (stdio) npm test # unit tests -- no network, no cost npm run testclient # simulate a customer end to end npm run wallet:new # generate a TESTNET wallet

Security in one paragraph

This server never holds a private key. Receiving crypto needs only a public address; only spending needs a key, and we only ever receive. Compromise the server and you get a cache and a ledger — you cannot get funds, because there is nothing to get. LLM spending is capped daily and checked before every call. Full detail: docs/SECURITY.md.

Honest status

Shipped, and used by nobody yet. As of 23 August 2026 the npm package is live and verified working from a cold install, and it has zero organic users. That is the honest state: the code works, the distribution has not started.

This is an unvalidated business. The measured facts:

• The entire independent x402 seller economy is ~$11,700/month across 14,128 registered services (measured 2026-08-07 — see MARKET_RESEARCH.md).
• The best independent operator makes ~$872/month.
• Most coding agents run inside a human's subscription and have no wallet.

So: expect free MCP usage to vastly exceed paid calls, and expect x402 revenue near zero in year one. The service is built so that outcome costs ~$1/month and still produces something genuinely useful.

The metric that matters is not revenue — it is calls per unique payer. Below 5 means tourism. Above 20 means a real business, even at tiny revenue.

Data Actors on Apify

The actors/ folder and src/jobs/ hold the source of CMM Research's paid data Actors. Each runs on Apify, with pay-per-result pricing and an MCP endpoint for AI agents:

Free samples on Kaggle: jobs, hospital jobs, funding rounds, Shopify stores.

Documentation

Data sources and ethics

All inputs are free, public, and unlicensed: the npm registry, PyPI, GitHub Releases, and OSV.dev.

We deliberately do not resell licensed data. The highest-earning independent x402 operators today proxy paid APIs (People Data Labs, Exa, Firecrawl) in probable breach of their terms. That is the one business model demonstrably working on x402, and we ruled it out.

We publish facts with short citations and links — never reproduced documentation.