Implementation BlueprintExecution layer

MCP Census MCP Server Vetting Gate (5-7 days)

A fixed-scope engagement that puts every MCP server in a client's agent stack through MCP Census evidence checks and leaves the client with a stamped PASS/REVIEW/BLOCK/UNKNOWN register plus a CI gate that keeps it current. Time: 5-7 days.

By InnovaAI ResearchPublished

AI Infrastructure

How do you implement it?

Blueprint

MCP Census MCP Server Vetting Gate (5-7 days)

A fixed-scope engagement that puts every MCP server in a client's agent stack through MCP Census evidence checks and leaves the client with a stamped PASS/REVIEW/BLOCK/UNKNOWN register plus a CI gate that keeps it current.

Prerequisites
  • Client inventory of every .mcp.json, CLAUDE.md, AGENTS.md, and .cursor/rules file across their agent repos
  • MCP Census account with an API key generated, and a decision on anonymous access (10 unique lookups/day per IP) versus keyed credits
  • Named client owner for the agent stack who can approve BLOCK decisions and server removals
  • Read access to the client's GitHub organization if the preflight gate will run via GitHub Actions
  • Written scope note stating MCP Census is a first gate, not a malware scanner or pen test, so the client does not expect either
Execution Timeline
  • 1.Create the MCP Census account and issue one API key per client environment so credit burn is attributable
  • 2.Map the client's agent surfaces: Claude Code, Codex, Cursor, VS Code, and any GitHub Actions pipelines that install servers
  • 3.Record the free tier baseline (1,000 credits each UTC month, no rollover) against the server count to size the plan
  • 1.Collect every .mcp.json config and run each through the /v1/audit endpoint for one verdict per entry
  • 2.Log the raw PASS, REVIEW, BLOCK, and UNKNOWN results with reasons and exit codes before any interpretation
  • 3.Flag entries that returned UNKNOWN so the client knows which servers have no public evidence trail at all
  • 1.Pull registry, GitHub, npm, and PyPI evidence for every REVIEW and BLOCK verdict and write the plain-language reason for each
  • 2.Separate BLOCK entries into remove-now versus needs-vendor-clarification so the client can act the same week
  • 3.Confirm hosted endpoint reachability for servers that claim a live endpoint
  • 1.Add the one-line Census rule to CLAUDE.md, AGENTS.md, and .cursor/rules so agents stamp servers before install
  • 2.Wire the GitHub Actions preflight step so pull requests that add a server fail on a BLOCK exit code
  • 3.Test the gate with a deliberately bad server entry to prove the exit code stops the build
  • 1.Subscribe webhook watches on the client's highest-risk servers so registry changes surface without a manual recheck
  • 2.Set up the change feed with a cursor so the client can see what shifted since the last audit
  • 3.Document the credit cost per watch and per audit call so the client can forecast monthly spend
  • 1.Deliver the stamped evidence artifacts and the verdict register as the client-facing risk report
  • 2.Walk the client owner through each BLOCK and REVIEW item and record their accept or remove decision
  • 3.Hand over the API key rotation and credit top-up procedure
  • 1.Run one full re-audit to confirm the CI gate and watches fire correctly in the client's own environment
  • 2.Close out with a one-page operating note covering who rechecks, how often, and what triggers an escalation
$1,800 fixed fee against $0 to $49/month in MCP Census subscription cost depending on server count and watch volume, plus $19/month if the client wants purchased credits that never expire5-7 days
ROI Logic

The client pays $1,800 for a 16-hour setup, while the tool itself runs $0 on the free tier (1,000 credits each UTC month) or $19 to $49 monthly for the builder and Team tiers, so the agency keeps the overwhelming share of the fee as delivery margin. Recurring revenue comes from a quarterly re-audit retainer priced against the client's actual credit burn rather than a flat markup, and the CI gate plus webhook watches make that retainer defensible because the client sees the gate working on every pull request.

Deliverables
  • Stamped MCP Census evidence artifacts for every server in the client's .mcp.json files
  • Verdict register listing PASS, REVIEW, BLOCK, and UNKNOWN with the reason and exit code for each entry
  • GitHub Actions preflight workflow that fails a pull request on a BLOCK verdict
  • CLAUDE.md, AGENTS.md, and .cursor/rules entries carrying the Census stamp rule
  • Webhook watch list and change-feed cursor setup documented for the client's own operators
Definition of Done

The client's agent repositories reject any new MCP server that returns a BLOCK verdict at pull request time, and every existing server carries a dated MCP Census verdict with its supporting evidence on file.

Open the MCP Census courseMCP Census decision recordBrowse the Academy

More on MCP Census

More for AI Infrastructure