Connect an agent to docir

Some agents run shell commands; some only call MCP tools. docir supports both, and the choice is about your client rather than about capability — the two expose the same vocabulary and cannot answer differently.

The CLI path: install the skill¶

For an agent that has a shell, install the instructions once per repository:

docir agent install                          # a Claude skill, plus an AGENTS.md block linking it
docir agent install --agent claude-writing   # add the document-writing rules
docir agent install --agent agents           # link the skills from AGENTS.md
docir agent update                           # refresh after upgrading docir

The generated files carry a version stamp, so update reports the transition. An AGENTS.md docir did not write is never rewritten — only docir's own marker block is.

The MCP path: one command¶

For a client that speaks only MCP, the server ships inside docir. There is no extra to install, because an agent that cannot run shell commands could not install one:

claude mcp add docir -- docir mcp serve

docir mcp serve speaks stdio, which is what an MCP client spawns. --transport http serves over HTTP instead, with --host and --port to bind it. uvx docir mcp serve runs the server without installing anything at all.

The tools are the commands¶

Every tool is one request through the same dispatcher the CLI crosses, so a tool and its command are the same code path. The names are prefixed — docir_context, docir_get, docir_add — because the CLI's verbs are generic and would collide in a client's tool list.

There is one tool per command, plus docir_schema for the thing an agent needs that is not a command. Results are trimmed exactly as the piped CLI output is, and reads return the same body-less skeletons.

Choosing between them¶

Install the skill when the agent has a shell: it teaches the whole workflow, including the parts that are judgement rather than invocation.

Reach for MCP when the client has no shell, or when you want docir available without the agent learning a CLI. Running both is fine — they are the same store through the same dispatcher, and neither knows the other exists.