Connect your agents
Teach coding agents the conciv CLI.
Agents use the conciv CLI directly. conciv tools is self-describing: agents run
conciv tools --help to discover every command, and the CLI finds your running dev server itself.
No MCP config, no addresses, nothing to maintain.
npx @conciv/cli@latest initinit scans for installed harnesses (claude, codex, gemini-cli, opencode, pi) and lists what it found in the plan
it prints before doing anything - ● claude for a harness it will teach, ○ codex (not found) for
one that is not installed. Pick Adjust at the prompt to choose the harnesses in a multiselect.
Approving the plan writes a marked conciv section into AGENTS.md (and CLAUDE.md if you keep one),
whether or not any harness was found; the section names the harnesses you picked in a
Set up for: line. Re-running init updates that section in place and reports it as
already wired when it is current.
claude is the one harness with a deeper integration: init generates the plugin files under
.conciv/claude-connect and installs them through claude's native plugin manager, scoped to the
project. If the claude CLI is not selected the step resolves as skipped: not selected, and if the
plugin manager fails init prints the two commands to run yourself.
init writes the marketplace, the plugin manifest, the MCP bridge, and a conciv skill that tells
claude to reach for the conciv CLI before any browser MCP tool, but no .mcp.json: the conciv
dev server writes that manifest when it connects to a claude session, so the plugin always points at
the address that is actually serving.
init prints the whole plan and asks once before it writes anything; --yes approves it up front and
--dry-run prints it and exits. It refuses a dirty git tree; commit first or pass --force.
Next: Usage to see what the agent can do.