Find the best tools for agents, and how we decide what to return

The directories that rank agent tools mostly rank them by GitHub stars, package downloads or how often a tool gets called through one proxy, and those numbers say what's popular without saying whether the tool answers at three in the morning, whether a model can make sense of its descriptions, how many tokens it costs before the first useful call or whether an agent can pay for it without a human ^[1]. The most-used documentation server in the MCP world ships a tool description of 2,006 characters that a community grader marked as failing ^[1]. We grade the things an agent runs into instead.

What we score

Every graded listing is scored on nine categories that add up to 100 points, weighted by what breaks agents in practice. Reliability carries 16, security and auth 14, schema and documentation 13, agent ergonomics 13, payments and pricing 10, maintenance and community 7 and transparency and trust 7, and performance and task success carry 10 each but are pending until our own probes and task suites run, so for now the other seven share the 100 ^[1]. Payments carry more than a payments category normally would, because a tool an agent can't pay for is a tool it can't use alone ^[1].

Research agents read each listing's public evidence, status pages and incident history, rate-limit and error docs, API references and OpenAPI files, MCP tool definitions in the source, pricing pages, terms, security pages, changelogs and repositories, and every listing carries at least four sources with the date each was read, a confidence and a list of what we couldn't check ^[1]. Five rules hold for every one of them. A score comes from evidence found and nothing gets points for what a vendor probably does, a thing we couldn't check isn't treated as absent without saying so, a vendor's claim is recorded as the vendor's claim, nothing is invented, and a fact on a listing that turns out wrong is corrected and named ^[1]. Up to 15 more points come off for incidents in the last twelve months that can be sourced, such as a security incident, a breaking change shipped without notice or a misleading claim ^[1].

The total maps to a grade from AA down to F, and BB or better is what we call agent-ready, meaning you can give the tool to an autonomous agent as long as you read the caveats on its page ^[1]. On 6 November one listing was AA, 22 were A and 184 were BB, so 207 of 1,491 were agent-ready ^[2].

What a search returns

A search for a job comes back in three groups, in this order, and every row says which group it's in.

An indexed listing has to clear a bar first, a hosted endpoint that answered tools/list in the last fortnight, at least 1,000 weekly npm or PyPI downloads or 300 GitHub stars, or a vendor's own verified domain, and of the 57,183 servers in the official MCP registry on 6 November, 2,000 cleared it ^[4]. The rest are still searchable, they just don't get our facts next to them.

The filters for what a company needs (an agent can start alone, no card, data in the EU, doesn't train on your data) answer yes, no or unknown for every listing, and the answer says how many it left out as each, because an agent can't tell "no" from "we haven't researched it" unless we tell it ^[3]. An agent can also pass its own weights, say security at 3 and payments at 0, and the search ranks by the same formula with its numbers in, which reproduces every published score to the decimal when every weight is 1 ^[3].

How letme picks one

letme is the part that gives an agent a single answer. Its published rule is the highest grade for the capability, then price, compared only between prices for the same job in one unit, then p95 latency, then the score, then the name, and each step only breaks the ties the one before it left ^[5]. A pick that's down right now, a major outage on its status page or a failed last probe, is passed over for the next one by the rule, and the answer says which one it skipped and why ^[5]. Nobody at our company can move a tool up that order by hand, and no vendor can pay to be picked ^[5].

What we don't do

Rankings are never for sale, and reports are paid while rankings aren't ^[1]. Our own products, letme and LocalGhost, are graded by the same checklist with a disclosure, two research agents grade them independently and a third reconciles them, and letme never picks either of them whatever their grades ^[1]. The research agents and the review panel run on Anthropic's Claude models, so Anthropic's listings are graded by the same checklist with every judgement call written down, and they carry a conflict label ^[1].

Why it matters

An agent pays for a wrong pick in tokens, in failed calls and in money, and it usually finds out halfway through a job. One of the most important bottlenecks for agents is the tooling, and that's why we exist. A grade with its reasons and sources next to it is something an agent can check rather than trust, and at CryptoCompare we learned that once the checklist is public, the people being graded start fixing the items on it ^[1].

Performance and task success are still pending until the probes and the task suites run, and the research on what a first call takes and what a company needs covered 107 of the 1,491 graded listings on 6 November ^[2], with more added since. If you think we've graded something wrong, file an issue with the evidence and we'll answer it in the open.

References

Read on 6 and 7 November 2026.

1. [1]Anchor Terminal, 2026. The benchmark, methodology 0.7. Source for the case against popularity rankings and the 2,006-character tool description, the nine categories and their weights, the evidence read and the five rules, negative events, the grade bands and agent-ready, the rules for our own products and for Anthropic's listings, the principle that reports are paid and rankings aren't, and what we learned at CryptoCompare. anchorterminal.com
2. [2]Anchor Terminal, 2026. The directory as JSON, read on 6 November 2026. Source for the 1,491 graded listings, the count at each grade, the 207 agent-ready and the 107 with setup or compliance research. anchorterminal.com
3. [3]Anchor Terminal, 2026. For agents, the search, its filters and weights. Source for the order of a search's answer, the yes, no and unknown answers of the requirement filters and the weights. anchorterminal.com
4. [4]Anchor Terminal, 2026. The index of listings built from public catalogues, read on 6 November 2026. Source for the bar an indexed listing clears, the 57,183 servers in the official MCP registry, the 2,000 that cleared it and the 2,222 indexed listings. anchorterminal.com
5. [5]Anchor Terminal, 2026. letme. Source for the published rule, passing over a pick that's down, and no manual overrides or paid placement. anchorterminal.com