Connect an agent over MCP
driga mcp serves read-only tools over stdio. Every response is the same
schema-2 envelope as --json, and reading never acknowledges inbox items. The
server cannot answer prompts, kill processes or remove anything; those actions
exist only inside Driga’s own Chat, behind approval cards.
Keep the installed driga binary on the client’s PATH, and run the
collector for history and merged fleet data. Without one, tools answer from a
labelled one-shot snapshot.
claude mcp add --scope user driga -- driga mcpIn ~/.codex/config.toml (or $CODEX_HOME/config.toml):
[mcp_servers.driga]command = "driga"args = ["mcp"]Machine, project and time scopes are tool arguments, not server flags. Pass
machine: "local" to read only this machine. driga mcp --snapshot forces
one-shot reads without history.
Every tool takes limit (1–500, default 100), which bounds the total entries
in a response including nested lists; truncated: true means some were left
out.
| Tool | CLI equivalent |
|---|---|
get_summary | driga summary |
list_agents, get_agent | driga agents, driga agents show |
get_flow | driga flow |
list_projects, inspect_project | driga projects, driga projects show |
list_inbox | driga inbox |
list_conflicts | Overlaps in Flow and Projects |
list_orphans | driga orphans |
get_machine, list_machines | driga machines show, driga machines |
get_activity | The web Activity page |
inspect_setup | driga doctor |
list_cleanup_candidates | driga cleanup list |
Inputs and descriptions for every tool are in the MCP tools reference, generated from the server itself.