The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the MCP Junction listing page.
A directory of Model Context Protocol servers, rebuilt nightly from public GitHub metadata. Roughly 1,800 servers in a hand-owned category taxonomy, also browsable by repository topic, published as a website, a downloadable dataset, and an MCP endpoint your agent can query directly.
mcpjunction.ai · Dataset · Categories · Licensing
Not affiliated with Anthropic or the Model Context Protocol project.
This repository is the whole thing: the pipeline that builds the dataset, the Astro site that renders it, and the Cloudflare Worker that serves it.
Three ways in, all public and all free to read.
Bulk download. The full directory, regenerated nightly with a generated_at timestamp.
mcp_servers.json — ~2.3 MB, 28 fields
per server plus a dataset envelope with counts and per-category totalsmcp_servers.csv — the same data as 16
flat columnsQuery it from an agent. /mcp is itself an MCP server — streamable HTTP, stateless
JSON-RPC over POST, protocol version 2025-06-18, no authentication.
| Tool | What it does |
|---|---|
search_servers | Free-text search, optionally filtered by category and language |
get_server | One server by id (owner--repo, e.g. microsoft--playwright-mcp) |
list_categories | Every category with active counts |
get_dataset_info | Counts, generated_at, source, and licensing terms |
GET /mcp returns 405 by design; POST your JSON-RPC messages. JSON-RPC batching is not
supported — it was removed in protocol 2025-06-18.
Read it as markdown. Every server and category page has a clean markdown twin with no navigation or styling to strip. Either negotiate for it:
…or just append .md to the canonical URL —
/servers/microsoft--playwright-mcp.md,
/categories/databases.md. Each HTML page
advertises its twin via <link rel="alternate" type="text/markdown">.
Machine-readable entry points are collected in
llms.txt.
Listing is automatic. There is no submission form, and none is planned — the pipeline runs
five GitHub search queries (topic:mcp-server, topic:model-context-protocol,
topic:mcp, and two name/description phrase searches), then keeps repositories that clear
a 2-star floor and match a Model Context Protocol name/topic heuristic. If your server is
on GitHub and meets that bar, it will appear on the next nightly run.
Corrections, removals, and licensing: licensing@mcpjunction.ai, acknowledged within three business days.
Open to read and reuse; deliberately narrow about what comes back.
Corrections are the most useful thing you can send. A wrong install hint, a stale
license, a repository that moved, a server sitting in the wrong category — open an issue or
email the address above with the server id (owner--repo) and what's wrong. One maintainer
cannot check 1,800 entries by hand, and the people who hit these find them first.
Pull requests are welcome for the software — src/, worker/, scripts/, docs/.
Bug fixes need no preamble. Open an issue first for anything that changes rendered output
or the dataset shape, because the constraints in the next section are load-bearing and easy
to undo by accident.
Three things don't take pull requests, for reasons rather than ceremony:
public/data/ — regenerated nightly, so the edit is gone within a day. If a server
is missing, the fix is upstream: tag the repository and the next run finds it.editorial/summaries.md — summaries are written in-house. A directory that other
people cite lives or dies on whether its prose is trustworthy, and reviewing a summary
properly costs about what writing one costs.categories.json and topics.json — every slug is a permanent public URL, and
category order is matching precedence rather than presentation. Propose one in an issue,
with the servers it would cover; adding a category is a decision, not a patch.The code is MIT and contributions to it are accepted on those terms. The dataset is separately licensed — see Licensing.
A directory other people cite has to fail safe. These are the constraints the code actually enforces:
security_reviewed, verified_badge,
sponsor_tier, editorial_notes, and editorial_summary are set by hand and only
carried forward by the pipeline — never written by it. A machine cannot grant a badge or
a security clearance here. These fields are currently unset across the entire dataset; a
written summary, when one exists, is a description and not a security review.status: archived_or_removed with a
delisted_at timestamp, then drop. Stable URLs, without unbounded growth.owner--repo name — a rename, or a released username claimed by someone else — the
editorial state is dropped rather than inherited.=, +, -, or @ are quote-prefixed
before publication. Repository descriptions are attacker-controllable: anyone can name a
GitHub repo.npx -y, no --yes. The registry package
sharing a repository's name may belong to someone else entirely, so hints are published
as unverified suggestions.Every night a GitHub Action runs pipeline.py, builds the site with Astro, generates the
sitemap, commits the refreshed dataset, deploys the Worker, and then verifies the live
result — licensing surface, /mcp handshake, crawler access rules, and markdown
negotiation — before it calls the run a success.
Locally:
The pipeline needs Python 3 and a GITHUB_TOKEN; deploys need Cloudflare credentials. See
docs/OPERATIONS.md.
The code is MIT. src/, worker/, scripts/, and the configuration are yours to
reuse — see LICENSE.
The dataset is not. public/data/** is excluded from the MIT grant and governed by the
RSL terms in public/license.xml: search indexing is free, agent and
inference-time use is free site-wide with attribution during the launch phase, and AI
training requires a license. Bulk /data/ retrieval carries stricter terms. Details at
mcpjunction.ai/licensing; see also NOTICE.
Citing us is easy and appreciated: via mcpjunction.ai.