Stork MCP
Officialby storkai
README.md
# Stork MCP
One connection, and your agent can search the web, read a page, look up
library docs, ask about a GitHub repo and set up other MCP servers. No keys,
nothing else to install.
Each job is done by the open MCP server that did best on Stork's own checks,
and every reply names it.
```
claude mcp add --transport http stork https://mcp.stork.ai/core
```
Then say what you want, **with Stork**:
| Say | Tool | Done by (today) |
| --- | --- | --- |
| "search with Stork for ..." | `stork_search` | Exa |
| "scrape this page with Stork" | `stork_scrape` | Firecrawl |
| "check the docs with Stork" | `stork_docs` | Context7, or the vendor's own docs for Microsoft, AWS and Cloudflare |
| "ask Stork about this repo" | `stork_repo` | DeepWiki |
| "set up Notion with Stork" | `stork_install` | Stork's directory of MCP servers |
| "find an MCP server for ..." | `stork_find_server` | Stork's directory of MCP servers |
Six tools, about 1,300 tokens of context. Free, 150 calls a day per caller.
Stork only calls tools that read: it will not send, pay for or delete
anything.
Other clients (Cursor, VS Code, Claude Desktop, Zed): add
`https://mcp.stork.ai/core` as a remote MCP server.
A client that only runs local servers can run Stork over stdio. It forwards
to the same hosted server:
```json
{
"mcpServers": {
"stork": { "command": "npx", "args": ["-y", "storkai", "--core"] }
}
}
```
or, from a clone of this repository, `node server/stork-mcp.mjs --core`.
## Claude Code plugin
The plugin connects the server and adds a skill, so "with Stork" works
without remembering tool names.
```
/plugin marketplace add storkai/stork-mcp
/plugin install stork@stork
```
## What is in this repository
| Folder | What it is |
| --- | --- |
| `plugins/stork/` | The Claude Code plugin: the server connection and the `with-stork` skill |
| `server/` | Stork as a local stdio server (one file, no dependencies) that forwards to the hosted one |
| `checker/` | The endpoint check Stork runs on every hosted MCP server in its directory, as one file you can run yourself |
| `question-sets/` | The questions Stork puts to every server in a category, and the script that asks them |
| `ROUTING.md` | Which provider each verb uses, and the numbers behind the choice |
| `METHOD.md` | What a check, a test call and a question set do and do not show |
The directory itself is at [stork.ai/mcp](https://www.stork.ai/mcp). The
service behind `mcp.stork.ai` is not in this repository; the parts that
decide what Stork says about other people's servers are.
## Check a server yourself
Node 18 or newer, no dependencies.
```
node checker/check.mjs https://mcp.exa.ai/mcp
```
```json
{
"outcome": "live",
"toolCount": 2,
"toolDefinitionTokens": 522,
"toolsFingerprint": "8522c5912e8cfadb",
"tools": ["web_search_exa", "web_fetch_exa"]
}
```
`outcome` is `live`, `sign_in`, `stream_only` or `no_answer`.
`toolDefinitionTokens` is what the server's tool definitions cost an agent in
context before it has done anything.
## Ask a server the question set
```
node question-sets/ask.mjs question-sets/web-search.json https://mcp.exa.ai/mcp
node question-sets/ask.mjs question-sets/web-search.json https://your.server/mcp --header "Authorization: Bearer $KEY"
```
Web search, 16 questions, 2026-11-17:
| Server | Key needed | Expected source found | Median time | Tokens per answer |
| --- | --- | --- | --- | --- |
| Parallel | yes | 16 of 16 | 2.5 s | about 4,400 |
| Exa | no | 15 | 0.5 s | about 3,300 |
| Tavily | yes | 14 | 1.0 s | about 1,900 |
This is a count, not a verdict: see `METHOD.md` for what "found" means and
why a server that returns more text has more chances. If you publish a
search server and think a question is unfair, or want a category added, open
a pull request against `question-sets/`.
## What Stork has found so far
From checking every hosted endpoint in the directory (shares, November 2026):
- 54% answered and listed their tools with no account, 36% asked for a
sign-in, 10% did not answer.
- Of the ones that answered with no account, a test call to one read-only
tool returned a result for 63%. For 13% the tool itself wanted a key or
account the handshake never mentioned. About 2% charge per call.
## License
MIT. See `LICENSE`.
This server cannot be deployed
Maintenance
ActivityMaintained
ResponsivenessNo recent community issues