Skip to content
Agentmart
Submit a tool

Changelog

Every user-facing change to Agent Mart, newest first. Follow along with the Atom feed.

AllDataTrustSearchSiteAPICLI

  1. · Site, Data

    Browse by use case and industry

    Listings are now tagged with what they are used for, which industries they serve and which roles they help, so you can browse by need instead of by tool type. Use-case pages cover 63 jobs, from browser automation to agent memory, and industry pages cover sectors such as financial services, healthcare and legal. Each page lists the top-ranked tools for that need across connectors, skills and other types, and some include an example stack. Pages with too few listings stay out of search engines until they fill up.

  2. · Site

    Share cards and a Cite this page button

    Listings, comparisons, best-of lists, guides and glossary terms now get their own share image, so links posted in chat or on social sites show the name and key facts instead of a generic card. Listing, comparison and glossary pages also have a "Cite this page" box with a ready-made citation and link, for anyone writing about a tool or quoting our data. Pages also carry structured data that search engines and AI assistants can read.

  3. · Site, Data

    Feeds, a changelog, New this week, and a newsletter you confirm

    You can now follow Agent Mart in a feed reader: an Atom feed of the newest listings at /feed.xml, one per listing type (for example /connectors/feed.xml), a security feed and a feed for this changelog. New this week lists every tool that joined the catalog in a given week, with type filters and a week picker. The weekly diff newsletter now uses double opt-in: you confirm your address before anything is sent, and every email has a one-click unsubscribe and a preferences page. The first issues are being prepared.

  4. · Site, Trust

    A leaderboard home page and rank on every listing

    The home page now leads with the leaderboard: an animated view of the current top 12, a ranked board with a tab for each kind of listing, and a section showing the real weights that move a tool up. Cards show their rank score, with their position when a list is sorted by rank. Each listing page opens with a rank panel: its position among visible listings of its type, its score and the five inputs behind it. Marketplace ratings now count only when backed by enough reviews, so a single five-star rating no longer moves a plugin.

  5. · Site

    Install guides and a glossary

    Sixty install guides walk through adding MCP servers and skills to Claude Code, Codex, Cursor, VS Code, Gemini CLI and other hosts, and connecting popular tools such as GitHub or Figma. Each guide names the host documentation it follows and the date it was checked, and 32 of them were run end to end with the real command-line tools in a clean environment. The glossary explains 54 terms, from MCP and A2A to tool poisoning and rug pulls, each with a link to a primary source.

  6. · Site

    Agent Market is now Agent Mart at agentmart.dev

    The site has a new name and a home of its own: Agent Mart, at https://agentmart.dev. Addresses that used the old /market prefix redirect permanently to the same page without it. The design changed with it: a new colour scheme, condensed headings, a wider layout and rounded cards across every page, with a compact header on phones. The listings themselves stayed the same.

  7. · Data, Trust

    See a server's tools before you install it

    Connector pages now have a Tools section listing the tools an MCP server offers, with each tool's description. For hosted servers we ask the server for its tool list; for local servers we read it from the source code. Tools the server marks as read-only, or as able to change or delete data, are labelled. Invisible or direction-changing characters in descriptions are shown as visible codes, so nothing in a tool description is hidden from you. We also check tool descriptions for hidden instructions aimed at your agent.

  8. · Trust

    Security scanning has started (grades come later)

    We have started scanning the source of connectors, skills and plugins for known vulnerable dependencies, leaked secrets, risky install scripts and obfuscated code. Scans read the exact published version and never install or run it. Results are not shown as grades yet: listings keep showing a dash until we publish the full method on the trust page. If a scan finds a real problem in someone's tool, we contact the maintainer privately first. Public pages will only ever show grades and counts, not exploit details.

  9. · Trust, Search

    Permissions panel and safer-by-default filters

    Listing pages now have a Permissions panel that sums up what a connector, skill or plugin asks for: its tools grouped into read, write and destructive, the secrets and environment variables it needs, the network hosts and file paths it touches, and whether it runs code on install. Cards show a Read-only chip where that applies. Three new filters, "Read-only tools only", "No secrets required" and "Up now", help you find tools that are easier to trust. The same filters work in the API.

  10. · Data

    Models with prices, context windows and a comparison table

    The new Models section brings in about 700 models from OpenRouter and Hugging Face. Each model page shows input and output prices per million tokens for every variant, how prices have changed, context window, license and whether weights are open. The comparison table can be sorted by price or context window, with the date each price was read and a link to its source. Filters cover provider, open weights, tool use, reasoning, image input, context size, price and free variants. Prices update daily.

  11. · API

    Search the catalog from your agent with our MCP server

    The catalog now has its own remote MCP server at https://agentmart.dev/mcp. Add it to Claude Code, Codex, Cursor or VS Code and your agent can search listings, read a listing's details, get the install config for its host, compare two listings, list categories and find alternatives. Every tool is read-only, and text written by publishers is cleaned and passed to your agent as data, not instructions. Setup steps for each client are on the developers page.

  12. · Data

    Infrastructure, templates and reference pages

    The catalog now covers twelve kinds of building blocks. Infrastructure, Observability and Identity list the services agents run on, with facts linked to each vendor's own pages. Templates gathers AGENTS.md and CLAUDE.md files, subagent definitions, Cursor rules and starter repositories, showing file text only when the license allows it. Reference covers protocols, specs and benchmarks such as MCP and A2A, each with links to the listings that use it.

  13. · Site

    Comparison and alternatives pages

    New comparison pages put two popular tools side by side: who should pick which, a table of stars, downloads, license, freshness and install options read live from the listings, and install steps for both. Alternatives pages list the closest options to a tool of the same type. Every page passes an automatic quality check before it is published, and pages that are too thin are kept out of search engines. Listing pages now link to the comparisons and alternatives they appear in.

  14. · Site

    Best-of lists and pages for each host

    Best-of pages rank the top MCP servers, skills and other building blocks for a topic, such as GitHub or agent memory, using the same score as search, and were reviewed for relevance. Host pages show what works with Claude Code, Codex, Cursor, VS Code and other clients, with install instructions checked against each host's own documentation. The footer now links to all of these.

  15. · Trust, Data

    Daily availability checks for hosted servers

    We now check every hosted MCP server once a day with a plain connection handshake, without credentials. Listing pages show an Availability panel with a 30-day strip, uptime and response time, and cards show a Down chip when a server failed its latest check. Servers that ask for sign-in count as alive. A server that stays unreachable for 7 days drops into low quality, and one that stays down for 30 days is hidden; it comes back automatically when it answers again.

  16. · API

    Public API v1 with an OpenAPI spec

    API v1 is live: search, listing details, install config for a given host, alternatives, comparisons, types, publishers and stats, with the same ranking and filters as the website. It is described by an OpenAPI 3.1 document and documented on the new developers page, with quickstarts in curl, JavaScript and Python. You can call it without a key (60 requests an hour, 500 a day); a free key, available on request for now, raises that to 10,000 a day. Errors share one format, and v0 keeps working with deprecation headers.

  17. · Data, Search

    Stars, downloads and new ways to sort

    Listings now show adoption numbers we read from the source: GitHub stars, plus downloads from npm, PyPI, Docker Hub, crates.io and NuGet where a package exists. Licenses are filled in from GitHub and the package registries when the listing itself did not say. You can sort any list by Most stars, Most downloads or Recently updated, next to Best match, Newest and A to Z. The numbers refresh every night.

  18. · Data

    Skills at scale, with install steps for each host

    Skills are now a full part of the catalog. We find SKILL.md files across public GitHub repositories and community lists, drop copies of the same skill, check each one against the skill spec and sort them into categories. Every skill page has install steps for Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot and other hosts that load skills. New filters narrow results to spec-valid skills, skills that ship scripts, skills that use the network, or open-source ones. New skills are picked up every night.

  19. · Data, Trust

    Spam, test entries and duplicates leave the default results

    Every listing now gets a quality tier. Test entries, spam and near-empty listings drop out of search, browse pages and the home page by default; choose "Include low-quality results" if you want to see everything. Duplicates point to the main listing for the same project instead of competing with it, and hidden pages explain why they are hidden with a link to ask for a correction. Servers published by the company behind the product carry an Official badge, and the new "Official only" filter shows just those.

  20. · Data

    Plugins from Claude Code, Codex, Gemini CLI and Open VSX

    The new Plugins section lists plugins and extensions for Claude Code, Codex, Gemini CLI and the Open VSX marketplace. Each plugin page shows what is inside it: commands, agents, skills, MCP servers and hooks, with links to the bundled listings we already track. Plugins that include hooks get a clear note, because hooks run code automatically when your agent does something. Filters find plugins that bring MCP servers, skills or hooks. Open VSX is refreshed weekly, the rest nightly.

  21. · Data

    The catalog now refreshes every night

    Listings no longer sit still between manual imports. A nightly job now pulls the latest servers from the official MCP Registry, merges them with what we already have and re-scores every listing, so new servers usually appear within a day and changed descriptions, versions and install details catch up on their own. When something disappears upstream we mark the listing as removed rather than deleting it. Anything we missed can still be sent through Submit.

  22. · Data

    Harnesses, agent builders and open-source agents

    Three new sections cover the rest of the agent stack. Harnesses lists frameworks and SDKs for building agents, Builders lists visual agent and workflow builders such as n8n and Dify, and Agents lists finished agents that do a job end to end. Each page has a fact sheet (languages, hosting, autonomy, protocol support, pricing where published) where every fact links to the page it came from. Filters cover language, autonomy and hosting. The lists started from hand-checked seeds and are refreshed weekly.

  23. · Trust, Data

    Fairer freshness grades and ranking

    Freshness grades now use the latest of three dates: the last commit (or the last commit to the listing's own folder in a larger repository), the last GitHub release and the last package publish. A means updated in the last 30 days, B 90, C 180, D a year, and F older or archived. Ranking changed too: adoption now blends relative and absolute popularity, so a 60k-star project is no longer tied with a 7k-star one, and a repository linked by many unrelated publishers only counts for its real owner. How it works is on the trust page.