Seven places that index agent tooling, on the same six criteria

These seven are not one market, and a single order over them would hide the thing that decides your answer. Three of them are businesses that run what they list: Smithery hosts the connection and charges for calls, Glama holds the largest catalogue on the grid and licenses it back to you with an attribution clause, and mcp.so sells the placement itself at $39 a submission and $399 a month for a slot. Two of them are hand kept files in public repositories: the punkpeye list is the broadest record of installable MCP servers and moves when a pull request is merged, and the e2b list is the only one here whose subject is the agents rather than the plumbing. One is the protocol project’s own registry, which verifies that a publisher owns the namespace they publish under and answers a public API with no key, and is still labelled preview. Agentwire is the last row and the smallest by three orders of magnitude, because it is the only one with no submission route at all: nothing can be filed, nothing can be bought, and an entry exists because two independent affirmations already existed elsewhere. So the question is not which of these is best. If you need coverage, the size of the catalogue is the criterion and the two largest rows win it outright. If you need to run the server, Smithery is the only row that does. If you need the data, read the API column and the licence attached to it. And if what you want is the shorter list of things somebody unconnected to the tool already vouched for, that is a different question, it produces a far smaller answer, and none of these tests what it lists, including this one.

Seven indexes, six criteria, the same six for every row. Each cell is the short answer. The same cell in that project’s own words is in its section below, and every source is listed at the foot of this page with the day it was read.

Where the entries come from. Whether the entries were supplied by the people who built the tools, found by a crawler, or copied from another list. It is what decides whether a listing is evidence of anything beyond somebody wanting to be listed.

How it stays current. An index of a fast moving category is worth what its last update is worth, and most of these do not print one.

Whether you can take the data. Whether the whole list is fetchable as data, what that costs, and what you have to do in return for it.

Nobody paid to be on this grid and nobody can. There is no sponsorship, no affiliate link and no submission route anywhere on this site. A figure a project does not publish is printed as not stated, never as a guess, and that is not a claim that no such figure exists.

This site is a row on its own grid, on the same six criteria at the same length, including the ones it loses. It is the smallest index here by three orders of magnitude, there is no way for the author of a tool to get it listed, and it does not test what it lists.

What is in it: Agent tooling of three kinds, and the smallest count on this grid. Its own scope page says "MCP servers, coding-agent harnesses, agent frameworks, and the tooling around them". The status page printed 149 entries on the day this was read, which is smaller than every other row here by three orders of magnitude in one case. That is the trade the no submission rule buys and it runs in both directions.

Where the entries come from: Curated lists other people publish, Hacker News, and a repository’s own metadata. Five signals, each of them somebody else’s judgement rather than ours: "A maintainer added it to a curated list they publish", a Hacker News thread above 40 points, created in the last 120 days and already past the star floor, past the absolute star floor, or two or more agentic topics in the repository’s own metadata.

How something gets listed: There is no route. Nothing can be submitted. No form, no pull request and no way for the author of a tool to add their own work. An entry appears when two of the five signals fired independently. Its own page: "Two independent signals is somebody else’s editorial judgement, arrived at twice, which is the entire product of a curation lane." The same page states that it does not test what it shares.

How it stays current: New entries daily, with the newest entry’s date printed. The status page answers whether the site is up, how many entries the index holds, and when the newest one was posted, which was 18 September 2026 on the day read. The digest goes out on the same cadence: "The newest entries, once a day."

Whether you can take the data: The whole index as plain text, with no key and no account. The llms.txt file lists every entry with its owner, its language and that owner’s own description of the project, and /api/search-index answered JSON to a plain request on the day read. Neither asks for a key, a sign in or an attribution clause.

What it costs: Free, and there is nothing to buy. Its own pricing page: "There is no account, no trial and no paid tier" and "Nothing here is a sponsored placement, and there is nowhere on this site to buy one." It is the only row on this grid where nothing can be paid for at all.

Pick it instead of this when. You want to connect an agent to a server rather than read about one. Smithery hosts the connection: it handles OAuth, credentials and sessions, and a published server gets a gateway URL an agent can call. Nothing else on this grid runs the thing it lists, and the free tier carries 50K remote procedure calls a month.

What is in it: Public MCP servers and skills, in a registry it also runs. Its API reference describes the listing endpoint as "Search and browse public MCP servers in the Smithery registry", and the same reference carries a separate set of endpoints for skills. The catalogue and the runtime are the same product, which is what separates this row from the lists.

Where the entries come from: The people who built each server, publishing it themselves. The publish guide is addressed to the server’s author: "Publish your MCP server on Smithery for distribution and analytics." There is no crawler and no editorial pass described anywhere in the documentation read.

How something gets listed: Publish it yourself, on their hosting or your own. Two routes on the same page. "Bring your own hosting" takes a public HTTPS URL and proxies to it, with streamable HTTP transport and OAuth support named as the requirements. The other uploads the server itself. Either way the person listing it is the person who owns it.

How it stays current: Not stated on the pages read. The documentation index describes publishing, releases and runtime logs, and states no cadence at which the catalogue itself is rebuilt or re-checked. That is not a claim that it is stale, only that the pages read did not carry it.

Whether you can take the data: A documented REST API over the registry. The reference publishes a "List all servers" endpoint on its platform API with full text and semantic search and filters for deployment status, verification and ownership. It is a product API with tokens rather than a public dump, and the same reference documents the token endpoints beside it.

What it costs: Free to 50K calls a month, then $10 a month. Its pricing page prints "Hobby" at "Free" with "50K" remote procedure calls a month and 3 namespaces, and "Pay as you Go" at "$10/mo" with 100K and 100, footnoted "RPCs consume your $10 credits, then $0.10 / 1K". What is priced is connecting to a server, not being listed as one. The page is drawn by client code, so this cell is re-read with a browser.

• Read it yourself at smithery.ai

Pick it instead of this when. You want coverage and you want it as data. Glama is the largest index on this grid by a wide margin and it publishes a documented directory API with cursor pagination at 100 requests a second, which is enough to mirror the whole catalogue in a few hundred calls. Nothing here comes close to it on breadth.

What is in it: Tens of thousands of MCP servers, plus remote connectors. The list printed 89,523 servers on the day read, and the reference separates the two kinds: a server is source code you can run, a connector is a running deployment with a URL, a transport and an authentication scheme. Only approved connectors reachable over streamable HTTP are listed.

Where the entries come from: Submissions, its own crawler, the official registry and an awesome list. Named on its own reference under Data Sources: community submissions through the servers page, "Automated discovery with manual curation", integration with the official MCP server registry, and the punkpeye awesome list. It is the only row here that states it copies from other rows on this grid.

How something gets listed: An Add Server control, or its crawler finds you. The list page carries an Add Server control, so an author can file their own work, and the reference says discovery also happens without anybody filing anything. At 89,523 entries most of the catalogue arrived the second way.

How it stays current: It prints its own last update time at the top of the list. The heading above the first row reads as a count and a timestamp together, and it said 2026-09-19 11:27 when this was read. It is the only row on this grid that publishes a freshness figure on the page a reader lands on.

Whether you can take the data: A directory API, with a key and an attribution licence. The reference says "Directory read endpoints require an API key" and rate limits every Glama API to 100 requests a second per address. It also says "This data is licensed, not public domain": every page that shows the data must carry a visible credit linking to Glama without nofollow, and every server shown must link to its own Glama listing. Attribution can be waived under a commercial licence.

What it costs: Not stated on the pages read. The servers list answered a plain request with no account on the day read. The pricing address served a sign up page carrying no figure, and the API reference prices neither the key nor the commercial licence it offers. The page read did not carry it.

• Read it yourself at glama.ai

Pick it instead of this when. You are looking across kinds rather than inside one. It carries MCP servers, remote servers, clients, agents and skills as separate directories under one search, so it answers what exists around the protocol rather than only which servers exist. It also runs on open source anybody can read: its site code is published as chatmcp/mcp-directory.

What is in it: MCP servers, remote servers, clients, agents and skills. Five directories under one search, listed across its own navigation, with a featured set on the front page. The front page describes the job as connecting AI apps to tools, data and automated workflows rather than as a catalogue of one kind of thing.

Where the entries come from: Submissions, with a featured set the site picks. The front page labels its top block "Featured servers" and captions it "Hand-picked, production-ready", so placement is editorial on top of whatever was submitted. Rows carry an added date rather than a checked date.

How something gets listed: A submit form, with a paid option at $39. The form takes a repository URL and a name and offers a "Paid submission" at "$39", a "one-time publishing fee", listed against "Publish immediately without review", a verified badge, featured and priority placement and a dofollow link. The form as served on the day read named no free route beside it. It is the only row on this grid where paying changes whether and how fast an entry appears.

How it stays current: Not stated on the page read. Rows print how long ago an entry was added, so the front page shows recency per entry. Neither the front page nor the submit page states a cadence at which the directory is rebuilt or its links re-checked.

Whether you can take the data: Not stated on the pages read. The front page and the submit page name no public API, no export and no feed. The site software is open source at chatmcp/mcp-directory, which is the code rather than the listings. The pages read did not carry it.

What it costs: Free to read, $39 to submit, $399 a month to be seen. Three sponsor tiers on its advertise page: "Platinum Sponsor" at "$1,299" a month site wide, "Gold Sponsor" at "$699" on listing and detail pages, "Silver Sponsor" at "$399" on detail pages only, billed monthly, with sponsors rotating within each slot. The $39 submission is separate and one off.

• Read it yourself at mcp.so

Pick it instead of this when. You are writing a client rather than reading a list. It is the only entry here that verifies the publisher controls the namespace it publishes under, and its API is public, paginated and free, which makes it the one an MCP client can build against without asking anybody for a key.

What is in it: Servers published by the people who maintain them. Its own README: "The MCP registry provides MCP clients with a list of MCP servers, like an app store for MCP servers." It is run by a working group of the Model Context Protocol project rather than by a company, and the README names its members.

Where the entries come from: The maintainer of each server, and nobody else. Publishing runs through the registry’s own publisher tooling and the thing it verifies is namespace ownership, that whoever publishes controls the name they publish under. That is a guarantee about identity rather than about quality, and how it compares with the other MCP listings is covered on the deep page.

How something gets listed: Publish it yourself, after proving the namespace is yours. The README leads with a publish link into the quickstart, and the registry rejects a publish whose namespace the caller cannot prove. There is no editorial step and no queue of pull requests in front of it.

How it stays current: A live API, and still labelled preview. The README records the preview launch on 8 September 2025 and an API freeze at v0.1 on 24 October 2025, with general availability still to come, and warns that "breaking changes or data resets may occur". A data reset is a different kind of staleness from an old commit and it is the sentence to read if you are building against it.

Whether you can take the data: A public API, paginated, with no key. Its servers endpoint answered a plain request on the day read and returned records carrying a published schema URL, a name, a description and the remote endpoints each server advertises. Nothing was signed in and nothing was asked for in return.

What it costs: Free, and nothing is sold. The repository’s LICENSE file records the project moving from the MIT License to Apache-2.0, with documentation under CC-BY-4.0. There is no hosted tier, no sponsorship slot and no submission fee anywhere in it.

• Read it yourself at github.com
• In depth here: what this listing checks, and who may add to it

Pick it instead of this when. The question is whether an MCP server exists for some API at all. It is the broadest hand kept list of installable servers, it costs nothing, it is one Markdown file you can read end to end or grep, and a missing server can be added by anybody the same afternoon.

What is in it: Installable MCP servers, categorised by hand. Its contributing guide sets the boundary: "This list is for servers with a public GitHub repository". A server that is only a hosted URL with no installable package is sent to a separate list, so this one is about code you run rather than endpoints you call.

Where the entries come from: Whoever opens a pull request, the server’s author included. The guide asks a contributor for the server name linked to its repository, a short description and the right category. What that checks is that the entry is filed correctly, which is what a hand kept list of this size can do. Who may add one, and what that makes a listing evidence of, is the subject of the deep page.

How something gets listed: Fork it, edit the README, open a pull request. Six numbered steps in the guide, ending at "Create a pull request". It also publishes a route for software: "If you are an automated agent, we have a streamlined process for merging agent PRs." A marker in the title opts in and the pull request is fast tracked, so a listed server may have been filed by an agent on its author’s behalf.

How it stays current: Whenever a maintainer merges a pull request. There is no schedule and there is no crawler. The list moves when somebody files an entry and somebody else merges it, which makes its freshness a function of the queue in front of it rather than of a job.

Whether you can take the data: One Markdown file, cloneable, with no key. The whole list is a README in a public repository, so it can be fetched raw, diffed between commits and read without an account. It carries no machine schema, so anything built on it is parsing prose.

What it costs: Free, MIT licensed. The repository page carries the licence as "MIT license". Nothing is sold, and there is no paid placement on it.

• Read it yourself at github.com
• In depth here: what this listing checks, and who may add to it

Pick it instead of this when. You are looking for the agents themselves rather than the plumbing under them. It is the only entry on this grid whose subject is AI assistants and agents as products, with a category, a description and a link set written out per entry, and a separate list is pointed at for the SDKs and frameworks around them.

What is in it: AI assistants and agents, described one by one. Its own boundary: "This list is only for AI assistants and agents." Each entry carries a category, a description and a set of links, and frameworks and SDKs are sent to a sibling list rather than mixed in. That is a narrower subject than this index, which carries the harnesses and the frameworks too.

Where the entries come from: Submissions and the maintainers’ own reading. It says what it is rather than claiming coverage: "The list is done according to our best knowledge, although definitely not comprehensive." The people who keep it are named on the page as the team behind e2b.

How something gets listed: A pull request, or a form. One line under "Have anything to add?": "Create a pull request or fill in this form", with alphabetical order and the correct category asked for. The form route is the only one on this grid that does not require a GitHub account.

How it stays current: Not stated on the page read. The README states no review cadence and prints no last checked date against its entries. Its own sentence about not being comprehensive is the closest the page comes to describing how it is kept.

Whether you can take the data: One Markdown file, plus a filterable web view. The list is a README in a public repository and the page also points at a web version of itself where the entries can be filtered by category and use case. As with the other list here, the data has no schema.

What it costs: No price on the page read. The README carries no price, no account and no paid placement. It links the maintainers’ own commercial product at the foot, which is a different thing from charging for the list, and the page read did not carry a figure either way.

• Read it yourself at github.com

Two pages on this site go further than a row on this grid, and neither of them ranks anything either.

• Where MCP servers get listed in 2026, and what each listing checks
• Agent framework and coding agent tool licenses in 2026
• How eligibility is decided here, and what this site does not do
• Where the entries come from, list by list

PulseMCP is not on this grid, and why is a fact worth printing. Answered 403 to a plain request and to a real browser on 2026-09-19, with the body "We offer API-based access to our data. Please contact hello@pulsemcp.com for more details." Nothing about it could be read off its own page, so it is not a row: an entry filled from anywhere else would be a claim about a company that nobody checked.

Prices and policies move, so every source is printed with the day it was read rather than one date at the foot of the page. A job re-fetches each of these URLs and fails when the literal string a figure came from is no longer on it. It also re-fetches the 12 sources quoted on the two pages above, because both of them have promised since they shipped that their figures can be re-checked rather than trusted, and until this job there was nothing that re-checked one. It ran on 2026-09-19 and 52 of its 52 checks passed.

Nothing on this page is a verdict on whether a tool works. None of the seven indexes tests what it lists, including this one. If an entry above misstates a policy you maintain, say so at hello@thecompound.tech and it gets corrected.