Home

Nakatomi Wiki

A headless, agent-native CRM. Claude, ChatGPT, Cursor, Perplexity, and any other MCP-capable agent can operate it directly — no human UI in the loop.

Start here

• Quickstart — get a local stack up in five minutes
• Architecture — visual tour of how the pieces fit
• Deployment — Railway, Docker Compose, Fly.io, bare VPS
• Authentication — API keys vs. user JWTs
• Troubleshooting — what to do when something's off

Using Nakatomi from an agent

• MCP-Tools — every tool on /mcp with a worked example
• REST-API — the full HTTP surface
• Webhooks — subscribing, HMAC verification, retries
• Agent-Patterns — see AgentLab for battle-tested recipes

Subsystems

• Memory-Connectors — DocDeploy, Supermemory, GBrain bridge
• Ingest — CSV, JSON, vCard, text → CRM rows
• Export-Import — portable JSON round-trip
• Custom-Fields — workspace-scoped field registry
• Rate-Limiting — per-API-key throttling
• Dashboard — optional local audit UI

Project

• FAQ
• Repository — source + issues
• ROADMAP.md — what's shipped, what's next
• AgentLab.md — agent deployment recipes
• ETHOS.md — the principles behind every decision

At a glance

%%{init: {"look": "handDrawn", "theme": "dark"}}%%
flowchart LR
    Agents[Claude / ChatGPT<br/>Cursor / Perplexity] --> Nakatomi
    CLI[curl / scripts] --> Nakatomi
    subgraph Nakatomi[Nakatomi]
        direction TB
        REST[REST API]
        MCP[MCP server]
        Worker[Webhook worker]
    end
    Nakatomi --> PG[(Postgres)]
    Nakatomi --> Storage[(Files)]
    Nakatomi -.->|optional| Memory[Memory connectors<br/>DocDeploy / Supermemory / GBrain]

Loading

Single Postgres is the source of truth. The MCP server and REST API run in the same FastAPI process and share a pool. The webhook worker is a background thread; memory connectors are pluggable adapters that mirror CRM mutations outbound and answer recall queries inbound.