multica-mcp

An MCP server (stdio, TypeScript) that wraps the multica CLI so an LLM chat can drive Multica directly: open issues, assign agents, follow progress, chain workflows.

Why this exists

The Multica CLI is the main tool. It covers everything this MCP covers, plus more, and it is the right fit for most workflows. Nothing here is meant to replace it.

This MCP is for one specific situation:

I find Claude Desktop (regular chat) is the best environment for brainstorming and architecture discussions, but I needed it to be able to open Multica issues, assign agents, follow progress, and chain sequential workflows directly from the conversation, without switching to a terminal. That is why I built this MCP.

If you spend your thinking time in an LLM chat (Claude Desktop, Codex Desktop, or any MCP-capable client), this lets the chat itself talk to Multica. If you live in the terminal, keep using the CLI.

Concrete use cases

• Brainstorm in chat, then hand off concrete pieces to Multica agents without leaving the conversation
• Run a sprint planning from the chat: draft issues, set priorities, assign agents, queue follow-ups
• Review agent results from chat: pull run messages, read comments, comment back, re-open or re-assign

Requirements

• multica installed on PATH and authenticated
• Multica daemon running
• Node.js 20+
• pnpm

Build

cd /path/to/multica-mcp && pnpm install && pnpm build

The bundle is emitted at dist/server.js.

Run

cd /path/to/multica-mcp && pnpm start

Logs

Tool calls are logged as JSONL with timestamp, tool, params, duration_ms, success, and error when relevant. The log file is resolved in this order:

1. MULTICA_MCP_LOG_PATH environment variable (absolute path)
2. $XDG_STATE_HOME/multica-mcp/calls.log
3. ~/.local/state/multica-mcp/calls.log

Configuration

| Env var | Purpose | | ---------------------- | --------------------------------------------------------- | | MULTICA_APP_URL | Base URL used when building issue and project URLs | | MULTICA_MCP_LOG_PATH | Absolute path override for the JSONL log | | XDG_STATE_HOME | Standard XDG base dir for the default log location |

When MULTICA_APP_URL is unset, the server reads multica config show and, if nothing matches, returns a null URL (no localhost fallback).

Tools

multica_list_agents

Returns [{ id, name, provider, model_hint, description }].

{}

multica_list_projects

Returns [{ id, name, description, issue_count }].

{}

multica_create_project

Returns { id, name, url }.

{
  "name": "Onboarding revamp",
  "description": "Agent coordination project for the onboarding revamp."
}

multica_create_issue

Returns { id, short_id, title, status, assignee, url }.

{
  "title": "Add a login flow",
  "description": "Detailed context in markdown...",
  "assignee": "claude-sonnet",
  "project": "Onboarding revamp",
  "priority": "medium",
  "cwd": "/path/to/repo"
}

multica_list_issues

Returns { items, state, total, offset, has_more, next_offset }. Use offset together with next_offset to paginate past the first 100.

{
  "status": ["todo", "in_progress"],
  "project": "Onboarding revamp",
  "limit": 20,
  "offset": 0,
  "sort": "updated_desc"
}

multica_get_issue

Returns { id, short_id, title, description, status, assignee, project, priority, comments, task }. Set include_comments: false to skip comments on heavily-threaded issues.

{ "issue_id": "ABC-12", "include_comments": true }

multica_list_comments

Paginated comment listing with limit, offset, since (RFC3339). Returns { id, parent_id, author, author_type, content, created_at } per comment.

{ "issue_id": "ABC-12", "since": "2026-04-01T00:00:00Z" }

multica_issue_runs

Lists all execution runs for an issue (status, timestamps, errors).

{ "issue_id": "ABC-12" }

multica_issue_run_messages

Streams the messages of a single run, with incremental fetch via since (sequence number).

{ "task_id": "<task-uuid>", "since": 0 }

multica_issue_search

Full-text search on issue titles and descriptions.

{ "query": "login redirect", "limit": 20 }

multica_add_comment

Returns { comment_id, created_at }.

{
  "issue_id": "ABC-12",
  "content": "Additional context for the agent."
}

multica_update_issue

Returns { id, short_id, title, status, assignee, priority, updated_at }.

{
  "issue_id": "ABC-12",
  "status": "in_review",
  "priority": "high"
}

multica_get_runtime_usage

Per-runtime { runtime_id, name, tokens_today, models_used_today, usage_30d }. Use window to scope: today | 30d | both (default).

{ "runtime_name": "Codex", "window": "today" }

multica_runtime_list

Lists runtimes (provider, mode, status, last_seen_at). Needed to obtain the runtime_id used by multica_agent_create.

{}

multica_agent_create

Creates a new agent bound to a runtime.

{
  "name": "claude-opus",
  "runtime_id": "<runtime-uuid>",
  "description": "Senior architect for design and refactoring.",
  "instructions": "Keep replies brief and evidence-based.",
  "custom_args": ["--model", "claude-opus-4-1"]
}

multica_agent_update

Updates agent name, description, instructions, runtime, args, concurrency, visibility, or status.

{
  "agent_id": "<uuid>",
  "status": "paused",
  "runtime_id": "<runtime-uuid>"
}

multica_workspace_members

Lists workspace members. Required to build @ mentions in comments.

{}

multica_attachment_download

Downloads an attachment to a local path. Returns { path, attachment_id }.

{ "attachment_id": "<uuid>", "output_dir": "./downloads" }

multica_list_autopilots

Returns the autopilots defined in the workspace.

{ "status": "active" }

multica_get_autopilot

Returns a single autopilot with its triggers.

{ "autopilot_id": "<uuid>" }

multica_create_autopilot

Creates an autopilot (scheduled/triggered agent automation).

{
  "title": "Daily triage",
  "agent": "claude-opus",
  "mode": "create_issue",
  "description": "Triage new issues every morning."
}

multica_update_autopilot

Updates an autopilot (title, description, status).

{ "autopilot_id": "<uuid>", "status": "paused" }

multica_trigger_autopilot

Runs an autopilot once, on demand.

{ "autopilot_id": "<uuid>" }

multica_delete_autopilot

Deletes an autopilot.

{ "autopilot_id": "<uuid>" }

multica_autopilot_trigger_add

Attaches a cron trigger to an existing autopilot.

{
  "autopilot_id": "<uuid>",
  "cron": "0 9 * * MON-FRI",
  "label": "weekday-morning",
  "timezone": "Europe/Paris"
}

multica_autopilot_trigger_update

Updates cron, label, timezone, or enabled state.

{ "autopilot_id": "<uuid>", "trigger_id": "<uuid>", "enabled": false }

multica_autopilot_trigger_delete

Deletes a single trigger from an autopilot.

{ "autopilot_id": "<uuid>", "trigger_id": "<uuid>" }

Smoke test

cd /path/to/multica-mcp && pnpm test

The smoke test boots the server over stdio, verifies initialize and tools/list, then exercises the tools against a temporary project and issue before cleaning up.

Install per client

The paths and commands below match the current public docs of each client and have been applied locally.

Claude Desktop

Local MCP servers for Claude Desktop are declared in claude_desktop_config.json. On macOS:

~/Library/Application Support/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "multica": {
      "command": "node",
      "args": ["/absolute/path/to/multica-mcp/dist/server.js"]
    }
  }
}

Restart Claude Desktop afterwards.

Claude Code

Claude Code exposes claude mcp add-json with --scope user:

claude mcp add-json --scope user multica '{"type":"stdio","command":"node","args":["/absolute/path/to/multica-mcp/dist/server.js"]}'

This entry is persisted in ~/.claude.json.

Check:

claude mcp get multica

Codex CLI

Codex CLI loads its configuration from ~/.codex/config.toml:

[mcp_servers.multica]
command = "node"
args = [ "/absolute/path/to/multica-mcp/dist/server.js" ]

Equivalent CLI command:

codex mcp add multica -- node /absolute/path/to/multica-mcp/dist/server.js

Check:

codex mcp get multica

Codex Desktop

The Codex app reuses the Codex CLI configuration. As long as ~/.codex/config.toml contains the multica entry, there is nothing else to wire up.

Known limitations

• Source of truth for the billed model is multica runtime usage, not the agent self-report
• Codex uses -c key=value, not --profile
• No persistent memory on the Multica side yet; a dedicated skill will come later
• cwd is not a native flag of multica issue create; the MCP injects it as a markdown hint
• If the Multica daemon is down, tools return a structured error: daemon down, run 'multica daemon start'

Advanced usage

For advanced usage patterns and best practices — routing matrix, safe issue creation, sequential chaining, PQR (dual reviewer), tool-call caps, token optimization, known limitations, and anti-patterns — see SKILL.md.

It is written for agents driving Multica through this MCP, but it reads fine as a human guide too.

Sources

• Anthropic Help Center, remote MCP connectors vs local servers declared in claude_desktop_config.json: https://support.claude.com/en/articles/11175166-get-started-with-custom-connectors-using-remote-mcp
• claude mcp --help and claude mcp add-json --help for the official Claude Code command
• OpenAI, "Introducing the Codex app", on the app reusing the Codex CLI configuration: https://openai.com/index/introducing-the-codex-app/
• codex mcp --help and codex mcp add --help for the official CLI syntax and persistence in ~/.codex/config.toml