@h1veframework/mcp

An MCP server that gives Claude Code / Cursor your project's memory, branch context, and workflow actions — from H1VE, the governance & memory layer for teams building with AI.

Teams that ship with AI lose context and drift from decisions as the codebase grows. H1VE keeps an eternal, per-project memory and an anti-drift engine; this MCP server exposes that to your AI editor: the current feature's spec and state (no copy-paste), the relevant project memory on demand (recall_memory), and the delivery actions — all authenticated by a personal token.

Part of the H1VE toolchain, alongside the CLI nf.

→ h1ve.org · app.h1ve.org

Setup (zero paste)

With the CLI installed, log in once:

nf login       # opens your browser → authorize → stores the credential locally

Then plug in the MCP server with no key at all — it reads the same credential nf login saved:

claude mcp add h1ve -s user -- npx -y @h1veframework/mcp

-s user makes it available in every project (use -s local inside a repo to scope it there). The project auto-resolves from your git remote, so there's no H1VE_API_KEY or H1VE_PROJECT_ID to set.

Confirm: claude mcp list should show h1ve: ✔ Connected, and /mcp inside Claude Code lists the tools.

The tools

| Tool | What it does | |---|---| | get_current_feature | State of the current branch's feature: stage, days active, blockers, sign-offs, AI declaration | | get_spec | The feature's spec (markdown) | | recall_memory | The project's relevant durable memory — decisions, gotchas, invariants — for what you're working on (by entity/query). Consult it before changing an area, so you don't contradict a past decision. | | move_feature_stage | Move the feature to another stage (rules validated on the server) | | create_blocker | Open a blocker on the feature (you become its owner) | | submit_ai_declaration | Submit the feature's AI declaration (owner dev only) | | start_feature | Start an assigned feature: record the slug and return the git switch -c to create the branch |

Reads (get_current_feature, get_spec, recall_memory) work with a PAT or a service key. Writes (move, blocker, ai-declaration, start) require a PAT — a service key returns SERVICE_CANNOT_WRITE.

Alternative: explicit env (CI / no CLI)

claude mcp add h1ve -s user \
  -e H1VE_API_URL=https://app.h1ve.org \
  -e H1VE_API_KEY=nf_pat_... \
  -- npx -y @h1veframework/mcp

H1VE_API_KEY / H1VE_PROJECT_ID are optional with nf login — use env only to force a key/project (CI).

Troubleshooting

| Symptom | Fix | |---|---| | Won't connect / 401 / 403 | Make sure H1VE_API_KEY is a PAT (nf_pat_…) from app.h1ve.org/api-tokens. SERVICE_CANNOT_WRITE = a read-only service key on a write tool → use a PAT. | | 404 running npx | npm propagation right after a publish — wait a few minutes. | | command not found: claude | Install Claude Code first. | | Tools don't appear | claude mcp list should say ✔ Connected; restart Claude Code to reload. |

Requires Node.js 18.18+. Speaks MCP over stdio; logs go to stderr (stdout is the protocol). The key travels in the x-api-key header and is never logged. Legacy env names NEXUS_FLOW_* are still accepted.

MIT · built with H1VE, by H1VE 🐝