MCP

The Model Context Protocol is how a coding agent works against your workspace instead of against whatever you last pasted into it. Dots ships an MCP server that exposes your Dotabases as tools the agent calls on its own — searching your corpus, reading a row, writing a new DOT, drawing an edge between two of them, walking your code graph — in the middle of the task it is already doing.

The server is not a second copy of the platform. Most tools are HTTP clients of the same API routes the CLI and web app call. Locally handled paths stay inside that catalog: batch dispatches its member tools in process, and a dry run can validate and preview a write without sending the mutation.

Key Features

• One governed catalog. Tools reuse the platform's route schemas, permission checks, and audit boundary instead of maintaining an MCP-only contract. The generated MCP Tools reference is the canonical catalog: use it for the current tool roster, exact parameters, permission atoms, and model-facing descriptions.
• Scoped to the credential, not the host. Every tool declares the permission atom it requires, and only the tools whose atom the connecting credential holds are ever registered.
• Fail-closed scope. When the scope read fails — a rejected credential, an unreachable daemon, a network fault — the connection advertises no scoped tools rather than falling open to the full write catalog.
• Read-only mode. dots mcp --read-only drops every write-classed tool from registration, and batch then refuses a write member at call time.
• Two transports. A host spawns dots mcp on your machine and speaks over stdio, or connects to the same server mounted on the network at /api/v1/mcp.
• Curated resources. dots://human-md, dots://dotabases, and dots://dotabase/{slug}/row/{humanId} attach to a conversation natively.
• Retrieved text is fenced. Everything a tool returns as free text is wrapped in a <vault-content> boundary carrying the instruction hierarchy the agent must honor.
• Every write is on the record. A write tool sends its own name in a dedicated header, and the API records one audit entry when the mutation actually changed state.

Get Started

• Connect An Agent Via MCP — the exact configuration for Claude Code, Claude Desktop, and Cursor. Start here.
• Quickstart: Agents — connect a host and scope what it may do, in one pass.
• Connect Your Agent — the model behind the connection: what the agent receives, and how much of your workspace it can touch.

Resources

• MCP Tools — the generated per-tool catalog: every parameter, the exact wording the model sees, and the permission atom each tool requires.
• Issue Scoped Agent Credentials — narrow a token to specific atoms and an expiry before you hand it to an agent.
• REST API Overview — the routes used by route-backed tools.
• CLI Overview — dots mcp is a CLI command, and it resolves its instance, credential, and company through the same precedence every other command uses.