Connect Your Agent

Any assistant that speaks the Model Context Protocol (MCP) connects to your workspace and works with your knowledge directly: it reads rows, runs search, and writes new DOTs and edges as it goes.

How The Connection Works

The dots CLI is the launcher; the MCP server it starts is a separate component. When a host runs the launcher, the server speaks MCP over standard input and output and the host attaches its tools to the model:

dots mcp

The launcher hands the server the endpoint, credential, and active company the CLI already resolved, so the agent acts against exactly the workspace you do from the command line. The host's own configuration repeats none of those three values.

Two transports run on the same server:

• Local (stdio). The host spawns the launcher on your machine and talks to it over stdin and stdout. This is the usual desktop setup.
• Streamable HTTP. The platform mounts the same server on the network at /api/v1/mcp, for a host that connects remotely instead of spawning a process.

Either way, every tool is an HTTP client of a real platform route rather than a direct database reach. An agent reaches through MCP what the CLI and the web app reach and nothing beyond it, and every write is audited on the route that serves it.

What The Agent Gets

Over the connection the agent receives a catalog of built-in tools covering search and row reads, Dotabase and outline listing, row and edge writes, saved views and dashboard widgets, code-graph traversal, and the Superpowers it can invoke. A batch meta-tool groups several calls into one: reads run in parallel, writes run in order, and a failed write halts the remaining writes, so the response names the member that stopped it rather than leaving a partial batch unexplained. The connection also carries curated resources, including your HUMAN.md and the Dotabase catalog. MCP Tools lists every tool and the parameters it takes.

How Much Access An Agent Has

Access follows the credential rather than the host. Every tool declares the permission atom it requires, and the server registers that tool only when the connecting credential holds the atom — so the catalog an agent lists is exactly the catalog it can use. The scope is read from the API when the connection opens, and a failed read fails closed to the ungated batch tool alone rather than advertising a write surface the credential may not hold.

Two levers narrow the connection further:

• A scoped token grants only the atoms you pick when you issue it, so one agent can read your corpus while another can also write to it.
• A read-only connection drops the write catalog while the read-classed tools stay registered — including the Superpower tools and batch, which then refuses any write member at call time. Read-only is a property of the launched server, so it constrains the agent even if the credential itself is broader.

Set It Up

Connect an Agent via MCP carries the exact configuration: the config file each host reads, the one-line registration commands, and the read-only and token options.