The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
This repo currently indexes 7,747 unique MCP server links from the category docs. The README stays lightweight while the full directory lives in docs/*.md, data/catalog.json, and the registry MCP server.
This repo is a community directory plus a hosted index. Every useful contribution makes the MCP Index easier for people and agents to search, compare, install, and verify.
Choose the path that matches what you want to do.
If you maintain an MCP server:
Share your public MCP profile from https://www.tensorblock.co/mcp so users can inspect metadata, install configs, source links, and badges without reading the raw markdown.
Improve your existing entry with install command, transport, auth requirements, supported clients, docs URL, license, endpoint, and tool details. Use the metadata issue form if you do not want to open a PR directly; when the profile id and structured values are clear, automation drafts a metadata sidecar PR for maintainer review.
Claim your TensorBlock MCP profile with the claim profile issue form. Claims show community maintainer ownership metadata and do not imply TensorBlock verification.
Add the TensorBlock MCP Index badge to your project README so users can jump from your repo to the indexed profile.
If you build MCP clients, agents, or developer tools:
Use the hosted API at https://mcp-index.tensorblock.co to search servers, fetch normalized profiles, list categories, or generate install-config previews.
Request another install target with the client config issue form. Include the client name, expected config shape, example config, and official docs. Clear requests generate a draft client-config spec PR so maintainers can review the target before implementation.
Help improve the config generator for Claude Desktop, Cursor, Codex, VS Code, and future clients.
If you want to improve the index itself:
Fix duplicate, stale, broken, or poorly categorized entries. Use the broken entry issue form when you want maintainers to triage it. Clear dead-link reports can generate direct cleanup PRs, while duplicate, stale, category, safety, or unclear reports generate draft investigation PRs so cleanup work can be reviewed and tracked.
Add missing metadata that makes search and install generation more accurate.
Propose verification signals, ranking improvements, or better category rules. The catalog health checker also runs on a schedule to flag duplicate primary links, missing GitHub repos, archived repos, and disabled repos as broken-entry reports.
Join the TensorBlock Discord to discuss roadmap work before opening a larger PR.
Issue forms are routed automatically. When you submit a server, metadata update, profile claim, client config request, or broken-entry report, the repo adds the right triage labels and posts the next steps so contributors and maintainers can keep the workflow moving. Server submissions with a clear category can generate draft docs PRs, structured metadata updates or profile claims can generate draft metadata PRs, clear client-config requests can generate draft spec PRs, and clear broken-entry reports can generate either direct cleanup PRs or draft investigation PRs depending on the report type.
New server submissions also get intake status labels so the queue is easier to review at scale:
needs-metadata means required fields, a valid project URL, or category routing still need contributor/maintainer input.
duplicate means the submitted project URL already appears in the catalog.
ready-for-pr means automation created or updated a draft docs PR for maintainer review.
automation-blocked means automation generated a branch but GitHub permissions blocked PR creation.
Maintainers can run MCP Add Server Intake Refresh from GitHub Actions to backfill or repair these labels across open server submissions. Run it with dry_run=true first to preview changes, then rerun with dry_run=false to apply them.
New server entries can be simple, but high-quality metadata makes the profile much more useful. The best entries answer:
What can an agent do with this server?
How does a user install or connect to it?
Does it use stdio, sse, or streamable-http?
Does it require an API key, OAuth, bearer token, or no auth?
Which MCP clients does it support?
Where are the setup docs, source repo, license, and public endpoint?
After a PR lands on main, the deploy workflow rebuilds the catalog and profiles. The hosted API and public profile pages refresh after the Railway deployment succeeds.
The scheduled catalog health check uses the generated catalog to open catalog-health issues for duplicate links and stale or unreachable GitHub repositories. Those issues feed back into the same broken-entry report workflow, so verified dead links can become direct cleanup PRs and ambiguous cleanup work can become reviewable investigation PRs instead of staying as loose maintainer notes.