MCP tools
Your coding agent talks to verifAIed through two MCP servers: a hosted one for coverage and verdicts, and a local one that records browser tests.
Connecting and signing in
In Claude Code, verifaied setup claude registers both servers. Neither holds a token: they use the login from verifaied login. Other agents connect with an API token you create under Settings → API tokens. A revoked or expired token stops working. Set up has the steps for each agent.
Without any credentials, a client can still connect to the hosted server, list its tools and call docs to read this documentation. Every other tool needs an account.
The verifaied server
The hosted server has 15 tools. They read what verifAIed knows about your repositories and branches, upload coverage and answer whether a branch is done.
setupThe first call on a fresh connection: what is set up for a repository and the steps that are left. Free.docsThis documentation, for your agent: the list of pages, one page, or a search. Also carries reference notes written for agents. Free.check_doneThe definition-of-done verdict for a branch — the stop condition for an agent loop.fix_branchEvery gap on a branch — failing tests, untested and partial functions — as one LLM-written prompt. Paid. scope="frontend" narrows it to React components grouped with their unexercised handlers and render branches.list_repositoriesYour connected repos by GitHub installation.list_analysesPer-branch analysis history for one repo.list_untested_functionsFunctions in an analysis filtered by coverage state.get_function_detailSource + uncovered lines for one function.get_baseline_promptA deterministic, ready-to-use test prompt for one function. Free and unlimited.upload_local_coveragePush coverage (pytest coverage.json or vitest coverage-final.json) + the current source from your dev loop — no CI required. Set coverage_kind: "browser" (CLI: --browser) to record the browser-session dimension instead of unit coverage.clear_manual_coverageReset the branch's accumulated manual-recording dimension — every function back to Unexercised — to start a fresh pass. Browser coverage has no clear; every producer replaces its own contribution.verify_uiJudge a screenshot of the running UI. Paid. Give it what you were trying to do and it answers pass / fail / unclear with reasoning; omit the intent for a rubric audit. The verdict is recorded, so check_done can hold the branch to it.get_generated_promptRequest and poll for an LLM-ready meta-prompt for a function.get_bulk_promptAggregate prompt for many functions at once.wait_for_promptPolling companion for the two generate-prompt tools.
The verifaied-recorder server
The recorder runs on your machine, next to your code and your running app. Your agent uses its 7 tools to record a browser test and prove it passes. See Record with your agent.
list_testsThe browser tests the branch already has, with each one's status and the tests it replays first.start_recordingOpen a recorder on the running app for a named test and return a snapshot of the page. One recording is open at a time.run_stepsPerform a batch of actions and checks on the page, add them to the test and return the page as it is afterwards.read_pageRead the page again as it stands right now.delete_stepRemove one step from the test being recorded.finish_recordingSave the test as a Playwright spec in the repository, then replay it with video. The replay has to pass.cancel_recordingEnd the recording without saving.
The setup prompt
The hosted server also offers a setup prompt. It asks your agent to run the setup tool for the repository and work through the steps that are left. In Claude Code, type:
/mcp__verifaied__setup
In another agent, ask it to run setup for the repository.