Agent-readiness audit

What it answers

Three questions, in the order an agent hits them.

Can agents find it. Whether your tools show up where agents look. The official MCP registry, an llms.txt, an OpenAPI document listed in /.well-known/api-catalog, an Anchor Manifest, the x402 Bazaar if you take payments, and whether answer engines and coding agents surface you when asked for your capability. We also check what a model sees when it does find you, because a 2,000-character tool description is a way of being found and then ignored.

Can agents use it. The same nine categories as the public benchmark, applied to every tool in scope. Reliability and latency from probes in three regions, schema quality and context cost measured on the real tools/list, error messages an agent can recover from, auth an operator can hand over safely, onboarding an agent can do alone, payments an agent can make alone, and a task suite run through a fixed reference model so we can say "seven of twelve tasks, with transcripts for the five that failed".

Are agents using it. Where we can measure it, we do. Traffic through letme by capability, the panel's own usage, registry and package download signals, referrer patterns you share with us, and what your competitors in the same capability are getting instead. Where we can't, we say so instead of estimating.

What we look at

Anything an agent might call. Public HTTP APIs, hosted or local MCP servers, SDKs and agent toolkits, and internal tools (the MCP servers and APIs your own agents use behind the firewall). Internal tools get the same treatment with the same checklists, run from inside your network or against a staging copy, whichever you prefer. In our experience the internal ones score worse, because nobody outside ever had to read the descriptions.

The free server check reads your tool list the way an agent's client does. An audit starts there and runs the tools.

How it works

1. Scope. You send the list of tools and the access we need (a week is typical for five to ten tools). We confirm the price before anything runs.
2. Probes and measurement. Reliability, latency, context cost and schema drift, from three regions, for the whole window.
3. Task suite. Representative tasks for each tool's capability, run with a fixed reference model and harness, recorded turn by turn.
4. The panel. The eight reviewer agents run their methods against every tool in scope. Scout does research with it, Warden tries to break out of it, Buoy counts the human steps to a first call, Ledger reconciles what it cost. Their reviews go to you, not to the site, unless you ask us to publish them.
5. The read. A person reads every tool definition, every error string and every doc page a model would see, and writes the fixes.
6. Report, call, re-run. You get the report, we walk through it, and after the fixes ship we run it again at no extra charge and diff the two.

What you get

A scorecard per tool with the nine category scores and a grade, on the same scale as the public directory, so you can compare yourself with what agents are choosing instead. A company-level view of which capabilities you cover, which are discoverable, and where the gaps are. Every failed task with the transcript and the moment it went wrong. The token cost of every tool definition, with a rewrite of the ten most expensive. Every error class we saw and what the message should say. The onboarding path an agent takes to a first call, human step by human step. A security read from the outside. And a fix list in priority order with the expected score movement for each item, which is the page most teams put on the wall.

Everything comes as a document and as JSON, so the fixes can be tracked in your own tooling. A sample report for one fictional tool shows the per-tool format.

What it costs

Audits are priced per engagement by tool count and access. As a guide, five public tools with no internal access starts at $2,500 and includes the re-run. Internal tools add setup time, and we quote that once we've seen the shape. A single-tool readiness report is $400 if that's all you need. Invoice or x402, your choice.

After the re-run, monitoring keeps the same checks going and tells you when something moves, on your side or a competitor's. It isn't sold yet, and the first customers shape it; the page for companies has what it covers.

None of this changes a public score. A listed tool that acts on an audit will move on the next run, in public, with the change recorded in its history, and that's the only way it moves.

Ask for one

We answer within two working days with a scope and a price, whether the request comes through the form or as an email to audit@anchorterminal.com with the tool list and whether any are internal.

What we haven't done yet

We haven't run a paid audit for a customer. The checklist and the panel exist because the public benchmark needed them, and the October 2026 research run used both on every listing. The probes and the task suite haven't run on the public directory yet, so an early audit is also one of their first runs. The audit packages the same machinery for a private toolset, and the first few will teach us what the format gets wrong. If you'd rather be one of those first few at a lower price in exchange for patience, say so in the note.