Reindent MCP servers

Two public, read-only MCP servers, built and run by Reindent:

Both are read-only by construction: the only way to register a tool in this codebase stamps it readOnlyHint: true, and there is no code path that issues a mutating request. Neither server holds a credential — each reads the same public data a browser reads.

Connect

Any MCP client that speaks Streamable HTTP:

{ "mcpServers": {
    "agenticlist": { "url": "https://mcp.agenticlist.ai/" },
    "reindent-skills": { "url": "https://skills.reindent.com/mcp" } } }

Or run either locally over stdio:

npm ci && npm run build
node dist/agenticlist/index.js     # stdio server
node dist/skills/index.js

Related MCP server: Agent Directory MCP Server

See it work

node dist/demo/exercise.js both connects a real MCP client to both servers over stdio and exercises every tool, including the failure paths (unknown slug, path traversal, empty query). node dist/demo/exercise.js both http does the same over the hosted transport on localhost and then attacks the guards with hostile requests. Transcripts: docs/DEMO-TRANSCRIPT.txt, docs/DEMO-TRANSCRIPT-HTTP.txt.

Layout

src/shared/       the scaffold both servers reuse — validation, rate limiting, caching,
                  audit logging, error envelopes, stdio + hosted transports
src/agenticlist/  S-42, the AgenticList catalog
src/skills/       S-41, the skills library (reads github.com/reindent/skills, so it cannot drift from the site)
src/demo/         the client harness
docs/             SECURITY-NOTES.md · demo transcripts

The hosted transport (src/shared/hosted.ts, src/shared/lambda.ts) is what the public endpoints run; the infrastructure that runs it is not part of this repository.

Security posture

docs/SECURITY-NOTES.md states what the protections are — and their limits — without implying any that do not exist. Short version: zod validation on every parameter (the slug pattern is the traversal/SSRF guard), edge rate limiting, an origin secret between Cloudflare and AWS so the raw function URLs are useless, hard upstream timeouts and response-size bounds, structured audit logs, no credentials, no PII. There is no authentication, because the data is public.

License

MIT