flipperai

Give your AI agents your team's context. flipperai is the Model Context Protocol connector for Flipper — your company brain. It exposes four tools (get_context, record_decision, search_memory, get_team_member) to any MCP-speaking agent (Claude Code, Cursor, Codex).

Install

npx flipperai install --workspace <workspace_id> --token <mcp_token>

This:

1. Writes the MCP stdio config to ~/.claude/claude_code_config.json (macOS/Linux) or %APPDATA%\Claude\claude_code_config.json (Windows).
2. Detects Cursor and appends a context block to .cursorrules (non-destructive).
3. Verifies the token + workspace against the Flipper API.

Get a workspace id and token from your Flipper dashboard (or, if you self-host, the mcp:token CLI).

How it works — it's a thin bridge, not the backend

flipperai serve is a small stdio MCP server that proxies each tool call to the hosted Flipper API over Authorization: Bearer <token>:

agent ⇄ (stdio MCP) ⇄ flipperai serve ⇄ (HTTPS + bearer) ⇄ Flipper API ⇄ knowledge graph

The graph, retrieval, Headroom compression, synthesis, billing, and rate-limit enforcement all run server-side. This package holds no database, no Redis, and no model keys — so it installs anywhere and carries nothing sensitive beyond your workspace token. Token validation, per-token rate limits (§7.1), and workspace isolation (§15.2) are enforced by the API, not here.

Configuration

serve reads its config from the env the installer writes:

| Env var | Required | Default | |---|---|---| | FLIPPER_WORKSPACE_ID | yes | — | | FLIPPER_TOKEN | yes | — | | FLIPPER_API_URL | no | https://api.flipperai.in |

Point at a self-hosted or local API with --api-url at install time (it's persisted into the config), or by setting FLIPPER_API_URL.

Maintainers — publishing

npm run build          # tsc → dist/
npm publish --access public

prepublishOnly runs the build; only dist/ (plus this README) is published.