bowmark-mcp — Bowmark MCP over stdio (Node)
Bowmark turns the interaction-gated web into a callable function library.
An agent reads the library (get_library), writes a short JavaScript script
against it, and sends it to run — Bowmark executes it on the live sites in a
sandbox and hands back the result. The canonical server is hosted, streamable
HTTP, no auth required: https://api.bowmark.ai/mcp.
This package is a thin stdio bridge to that hosted server, for MCP hosts
whose client only speaks stdio, in Node-flavored environments. Tool schemas,
descriptions, and results pass through verbatim — the hosted server stays the
single source of truth; nothing is reimplemented here. (Python-flavored
environments: the same bridge exists on PyPI as bowmark-mcp — uvx bowmark-mcp.)
mcp-name: ai.bowmark/bowmark
Use
npx bowmark-mcp
Any MCP host config:
{ "mcpServers": { "bowmark": { "command": "npx", "args": ["bowmark-mcp"] } } }
If your host speaks streamable HTTP, skip this bridge and connect directly to
https://api.bowmark.ai/mcp.
Environment
BOWMARK_MCP_URL | Target MCP URL. Default https://api.bowmark.ai/mcp/npm (the /npm destination attributes the install to this bridge). Point at http://localhost:3001/mcp for a local Bowmark API. |
BOWMARK_API_KEY | Optional. Forwarded as X-Bowmark-Key; a free key (bowmark.ai dashboard) lifts the anonymous per-IP daily cap to your plan budget. |
Design notes (repo-internal)
- One remote session per request, retried once. The hosted MCP is
stateless, so a fresh connection is semantically identical and its cost (one
initialize round-trip) is noise next to a
run that drives real browsers — and it
sidesteps long-lived-connection failure modes without reconnect bookkeeping.
Mirrors packages/bowmark-mcp/python (the PyPI bridge) exactly.
- The real host's name is relayed upstream. The api tailors its operating
guidance per host and detects the host from
clientInfo on the handshake —
but through a bridge that names the BRIDGE, so every install here would read
the platform-neutral text. So the host's own clientInfo.name (from OUR stdio
handshake) is forwarded as X-Bowmark-Client on every proxied request, and the
api ranks that above its own handshake. Best-effort: an unknown or missing name
simply falls back to the neutral text, never an error.
- Pass-through only. No tool logic lives here; the agent-surfaces sync
rule in the root CLAUDE.md is unaffected because descriptions/schemas ride
through from
apps/api/src/routes/mcp.ts. callTool results are returned
verbatim (content, structuredContent, isError).
- The
/npm destination is registered in
apps/api/src/mcp-destinations.ts + mcp-registry/sources.json (npm stdio
bridge channel; the PyPI bridge is /pypi). It carries the channel and
deliberately pins NO platform, because the relayed X-Bowmark-Client above is
a better answer than any pin. The old ?s=n query form still resolves for
installs already in the wild. Adding a destination:
.claude/rules/mcp-destinations.md.
- npm ownership verification needs BOTH markers — the registry checks the
published package.json, not just the README. The
mcp-name: ai.bowmark/bowmark line above and the "mcpName": "ai.bowmark/bowmark"
field in package.json are both load-bearing; don't remove either. Learned
the hard way 2026-07-04: release-mcp.yml's registry republish failed for
five straight api releases (server returned status 400: ... "NPM package 'bowmark-mcp' is missing required 'mcpName' field") because only the
README marker was present — the package.json field is what mcp-publisher publish actually validates against the live npm package.
- Versioning is manual (thin bridge, not the api): bump
package.json
when it changes. Not wired into release-please; private is deliberately
absent so npm publish works — release-please doesn't manage this package.
Tests
Network-free unit tests live in the monorepo suite:
tests/unit/mcp-stdio-node.test.ts (pnpm test:unit). The remote hop is
injected at the callRemote/buildServer seams.
Publishing to npm
Live since 2026-07-04 (published by CI + cold-verified via
npx -y bowmark-mcp against prod). To ship a version: bump version in
package.json (and the two version literals in src/bridge.ts) and
merge. The actual publish does NOT run in this monorepo's CI — this repo
is private, and npm Trusted Publishing (OIDC) validates against
package.json's repository.url, which (correctly) points at the public
mirror github.com/bowmark-ai/mcp, not Metroxe/bowmark. So merging here
only lands the version bump; release-bowmark-mcp.yml mirror-syncs it to
bowmark-ai/mcp, and THAT repo's own .github/workflows/publish.yml
(source-controlled at
packages/bowmark-mcp/.github/workflows/publish.yml
in this monorepo, mirrored in like any other file) does the real npm publish. Auth is npm Trusted Publishing (OIDC) — no token, configured
on npmjs.com (package Settings → Trusted Publisher → GitHub Actions, repo
bowmark-ai/mcp, workflow publish.yml). Publishing from the public
mirror also auto-generates provenance attestations, which npm only produces
for public-repo publishes — impossible from this private monorepo
regardless of repository.url. Not release-please; the bump IS the release
action.
mcp-registry/server.json carries the matching npm packages entry (landed
after the first publish per the mcp-name ordering rule), and the website
stdio tab's Node step points at npx bowmark-mcp.