replen MCPサーバー

@replen/mcp

Smarter AI Development workflows. MCP server that brings replen - the AI that asks "can we do this better?" on your codebase - inside Claude Code / Codex / any MCP host.

While your AI coding tool waits for prompts, replen reads your code against the ecosystem every morning. This MCP lets the agent act on the results without leaving chat:

• "what new tools shipped this week that fit my project?"
• "is github.com/owner/repo worth integrating into the codebase I have open?"
• "open handoff PRs for everything I starred"

What it exposes

replen ships 15 tools. Three moving parts sit behind them: Brainstem matches, Watchtower watches, Atlas remembers. The tools group under Brainstem, Atlas, and onboarding.

Brainstem (the matching loop, in-session on your own subscription tokens)

Tool	Returns
replen_match	The cwd repo's candidate inventory: repo metadata + cosine scores + whyShortlisted lines, everything Watchtower surfaced and Brainstem scored against this codebase's capabilities
replen_record_triage	Records the agent's per-candidate verdict (adopt / port / cherry-pick / clean-room / upgrade / skip / defer) with score, effort, and reasoning
replen_state	Records the user's action on a candidate: star / hide / handoff / surfaced
replen_capture_insight	Stores a portfolio insight from triage: a transferable lesson or a sharpened boundary

Atlas (the knowledge graph + memory)

Tool	Returns
replen_leaps	Non-obvious cross-project / adjacency / cross-user connections from Atlas, each with a via path
replen_recall	Memory across your whole portfolio: past verdicts, grounded reports, capabilities, and notes for a query
replen_cart	Pulls a saved or built-in Atlas Cart's rows in-session (Blind spots, Triage board, Keystones, Brought in, Stale deferrals, By domain, or your saved carts)
replen_queue	The awareness-to-action queue: list / add / done / dismiss items carried over from the Brief
replen_handoff	Opens a handoff PR in the matched project's repo

Onboarding (grounding a project)

Tool	Returns
replen_onboard_state	Per-repo grounding state across the portfolio, the cheap pre-flight for /replen-onboard
replen_set_capabilities	Writes a project's grounded capabilities + report (read locally by the agent, never uploaded as raw files)
replen_set_tags	Writes a project's ranked domain tag cloud
replen_set_versions	Writes a project's pinned direct-dependency + runtime versions
replen_set_product	Groups sibling repos under one product

Plus replen_help, the tool index.

The core tool is replen_match: it returns the cwd repo's candidate inventory (metadata + cosine + whyShortlisted) with no LLM call on our side, so the host agent reasons about fit against your actual open codebase in context, then records each call via replen_record_triage. Atlas now has two halves: the graph/explore view and Carts, browsable filterable views over your decision graph that replen_cart reads in-session.

Install (one-liner)

Get your token from your replen /settings page → "Connect Claude Code", then run:

npx -y @replen/mcp setup --token=ing_xxxxxxxx --base=https://app.replen.dev

That writes the MCP entry into ~/.claude.json (with a backup of the original) and is fully idempotent - re-run it any time you rotate the token.

Restart Claude Code to pick up the new server.

What it actually does under the hood

• Adds mcpServers.replen to your Claude Code config
• command: "replen-mcp" - resolved via npx on each host launch, or npm i -g @replen/mcp for a slightly faster startup
• env: DIGEST_BASE_URL + DIGEST_TOKEN
• Existing MCP servers (firebase, neon, etc.) are preserved untouched
• A .bak of your config is saved next to it for one-step recovery

If you'd rather hand-edit your config, the block to add is:

{
  "mcpServers": {
    "replen": {
      "command": "replen-mcp",
      "env": {
        "DIGEST_BASE_URL": "https://app.replen.dev",
        "DIGEST_TOKEN": "ing_xxxxxxxx"
      }
    }
  }
}

Other commands

replen-mcp --version     # print version
replen-mcp --help        # show available subcommands
replen-mcp               # run as stdio MCP server (your host spawns this; you usually don't run it directly)

Example session

You: anything new today for my project X?
Agent: [calls replen_match({repo: "you/my-project-x"})]
       2 candidates for this repo, top one is roboflow/supervision - 38k★ - MIT -
       cosine 0.71 · whyShortlisted: fills your "object tracking" capability.
       Want me to triage them against your current codebase?

You: yes
Agent: [reads the candidate README + greps your src/ for related code]
       supervision drops in for your hand-rolled annotation utilities and
       deletes your ByteTrack reimplementation. Strong fit.
       [calls replen_record_triage({repo: "roboflow/supervision", verdict: "adopt", ...})]
       Recorded. Want a handoff PR?

You: yes
Agent: [calls replen_state({repo: "roboflow/supervision", status: "handed_off", ...})]
       [calls replen_handoff({matchId: 96})]
       PR opened: github.com/you/my-project-x/pull/142

Costs

• Read-only tools (replen_match / replen_cart / replen_recall / replen_leaps) cost nothing - pure JSON shuttle.
• Triage (the per-candidate verdict + writeup) is the agent's own reasoning, paid through your Claude subscription inside the session, never on our side. Replen stores no LLM API key and makes no per-candidate model call; the only server-side model work is cheap embeddings + a catalogue classifier on Replen's own account.
• replen_handoff is one GitHub write call. No LLM on our side.

No additional replen-side cost vs the web dashboard - the MCP server queries the same API as app.replen.dev.

License

Apache-2.0 - see LICENSE. Permissive open source: free to use, modify, and redistribute for any purpose, with an explicit patent grant. "Replen" and the other named surfaces remain trademarks (Apache §6).