Tiltmeter

Tiltmeter

An early-warning system for the third-party APIs your AI agents depend on, with the fix as a pull request. · tiltmeter.dev

Volcanologists bolt tiltmeters to the ground to sense tiny deformations long before an eruption. Tiltmeter does the same for the APIs under your agents. It learns the shape of their JSON responses from real traffic, notices when a vendor renames a field, starts sending null or turns a number into a string, and tells you, your CI and the agents themselves before anything goes on tilt. Then tiltmeter fix hands the change to an agent that adapts the code reading that API and opens the pull request, leaving what it cannot know for a human.

Typed code crashes loudly when an API changes. Agents don't. They read the new response, fill the gaps with guesses, or retry until the budget is gone.

Highlights

• Agents can check before they act. An MCP server lets Claude Code, Cursor, LangChain or any MCP client call has_drifted before trusting an API, and get_schema to see which fields to expect.
• No spec required. Tiltmeter learns from real responses, so it works with any JSON API, documented or not.
• Breaking or harmless, at a glance. Every change is classified as breaking, warning or info, and a removed field next to a similar new one is flagged as a probable rename.
• Explanations on demand. With an Anthropic API key, Claude explains what changed, the likely cause and what to do about it.
• Fixes, too. tiltmeter fix hands the drift to an agent that adapts the code reading that API, shows you the diff and opens a pull request. Use the built-in LangChain Deep Agent, or your own Claude Code or Codex login.
• A CI gate with alerts. Exit codes fail the build on breaking drift. Slack and webhook alerts take one line of config.
• Local and private. One SQLite file and no account. Credentials are expanded at request time and never stored. Response bodies are not stored or sent to the LLM, only their shape.

Quickstart

pip install tiltmeter        # or: uv tool install tiltmeter
tiltmeter init               # writes tiltmeter.yaml with a public example API
tiltmeter check              # the first check records the baseline

Already calling an API from a web app? Export a HAR from your browser's DevTools (Network tab, "Export HAR") and let Tiltmeter write the config from the JSON calls the page made. Tokens and cookies become ${VAR} placeholders:

tiltmeter init --from-har session.har --host api.example.com

Run tiltmeter check again from cron or CI, or leave tiltmeter watch running. This is what you get when a vendor renames phone and turns balance into a string:

$ tiltmeter check
✗ billing_customers  200 · 13 ms  BREAKING · 3 changes
    breaking  type_changed   $.data[*].balance       integer → string
    breaking  field_removed  $.data[*].phone         string → (absent)  possibly renamed to $.data[*].phone_number
    info      field_added    $.data[*].phone_number  (absent) → string  possibly replaces $.data[*].phone
1 endpoint checked · 1 breaking
$ echo $?
2

The getting started guide runs that exact scenario against a fake API on your machine in about five minutes.

Give your agents a tiltmeter

claude mcp add tiltmeter -- tiltmeter serve --config /absolute/path/to/tiltmeter.yaml

Then add a line like this to your agent's instructions:

Before you rely on a third-party API, call has_drifted for it. If it drifted, read the summary and adapt before trusting the response.

Setups for Cursor, Claude Desktop and LangChain agents are in the agents guide.

Let it fix your code

When a check finds drift, tiltmeter fix hands it to an agent that finds the code reading that API, adapts it, shows you the diff and opens a pull request:

$ pip install "tiltmeter[fix]"            # or plain tiltmeter, then --agent claude-code or codex
$ tiltmeter fix 2 --repo ~/code/billing-service
✗ billing_customer  #2  200 · 10 ms  BREAKING · 3 changes
Working on a new branch, tiltmeter/fix-billing-customer-2, from main at 162dadd.
3 files changed  +5 -5
diff --git a/billing/client.py b/billing/client.py
-    return f"{customer['name']} <{customer['phone']}>"
+    return f"{customer['name']} <{customer['phone_number']}>"
...
Commit, push and open a pull request? [y/N]: y
✓ opened https://github.com/acme/billing-service/pull/41

• Your checkout is never touched. The agent works in a temporary git worktree on a new branch, and nothing is committed until you approve the diff or pass --yes.
• It stays inside the worktree. No network access, and only tracked files are there to read; .env files, keys and .git are blocked. The built-in agent and Claude Code get file tools only. Codex keeps its shell, locked in its OS sandbox under the same limits.
• It says what it could not do. Tiltmeter sees types, not values. When a number turns into a string, the agent updates the types and leaves the format question to a human in the pull request instead of guessing a conversion.
• Bring your own agent. Already use Claude Code or Codex? tiltmeter fix 2 --agent claude-code or --agent codex runs on your own login, with no API key and no extra install.

--dry-run shows the diff and throws it away, and --no-pr commits to a local branch. The fix guide covers the flags, what the model sees and how to review the result.

What counts as drift

Override any of them globally or per endpoint with severity_overrides.

Use it in CI

On GitHub, the Tiltmeter Action keeps the baseline in the Actions cache, writes the changes to the job summary and fails the job on drift:

- uses: tiltmeter-dev/tiltmeter-action@v1
  env:
    STRIPE_API_KEY: ${{ secrets.STRIPE_API_KEY }}

Anywhere else, keep .tiltmeter/ between runs. The CI guide has ready-to-copy jobs for GitHub Actions, GitLab CI and cron.

How it works

1. Observe. A check calls the endpoint and turns the JSON body into a shape: for every path, which JSON types appeared, how many objects contained each field and, for fields that look categorical (active, past_due, USD), which values were seen.
2. Compare. The baseline is the merge of the shapes from the last 20 successful checks. Tiltmeter reports only what contradicts it: a missing optional field is fine, a field that was in every object and is now gone is not.
3. Report once. The new shape joins the baseline, so the check that sees a change reports it, and from then on it is the new normal.

graph TD;
  start([start]) --> observe
  observe -. fetch failed .-> persist
  observe --> load_baseline --> compare
  compare -. drift worth explaining .-> explain --> persist
  compare --> persist
  persist -. a notifier wants it .-> notify --> finish([end])
  persist --> finish

The pipeline is a small LangGraph graph that only sequences the steps. The diff itself is plain, deterministic Python with no LLM in the loop. How it works covers the details.

Configuration

endpoints:
  - id: stripe_customers
    url: https://api.stripe.com/v1/customers
    params:
      limit: 10
    headers:
      Authorization: Bearer ${STRIPE_API_KEY}   # expanded from the environment at request time
    ignore_paths:
      - $.data[*].metadata                      # free-form, changes all the time

notify:
  - type: slack
    webhook_url: ${SLACK_WEBHOOK_URL}

Every option is described in the configuration reference, and examples/ has a config you can run as is.

How it compares

• OpenAPI diff tools such as oasdiff compare two versions of a spec. They are the right tool when the vendor publishes and maintains one, and they cannot see changes the vendor did not document.
• Contract testing such as Pact needs the provider to run your contracts, which third-party vendors don't do.
• Uptime and synthetic monitors tell you an API is up. You write and maintain the response assertions yourself.
• Tiltmeter learns the contract from real responses, so it needs nothing from the vendor and catches undocumented changes. It also hands what it learns to your agents.

Privacy and security

• ${VAR} placeholders are expanded right before a request or notification is sent. The values are never written to disk, sent to the LLM or kept in the LangGraph state, and they are scrubbed from error messages.
• The history stores shapes, status codes, latencies and changes, never response bodies unless you set store_samples: true.
• Agents can only check the endpoints in your config. They cannot make Tiltmeter call an arbitrary URL.
• tiltmeter fix sends the change list and the files the agent opens to Anthropic. It never sees response values, ${VAR} values or untracked files such as .env, and paths matching fix.ignore are blocked.

SECURITY.md has the full threat model and how to report a vulnerability.

Roadmap

• Fixes in CI and over MCP. Run tiltmeter fix from the GitHub Action when a scheduled check finds drift, and let agents request a fix through an MCP tool.
• Provider mode. API vendors announce a change once and every customer gets the pull request.
• Spec and changelog awareness. Diff vendor OpenAPI specs and changelogs next to observed responses.
• Hosted checks and a team dashboard.

Documentation

• Getting started: a five-minute tour on your machine.
• Configuration reference: every option, placeholders and ignore patterns.
• Agents and MCP: Claude Code, Cursor, Claude Desktop and LangChain.
• CI and scheduling: GitHub Actions, GitLab CI and cron.
• Fixing code: tiltmeter fix, the agent and how to review its pull requests.
• How it works: shapes, baselines, the diff rules and the pipeline.
• FAQ and troubleshooting.

Contributing

Bug reports, ideas and pull requests are welcome. CONTRIBUTING.md explains how to set up the project and what a good change looks like. Everyone taking part is expected to follow the code of conduct.

License

MIT