Phase 0Done
Foundations
Monorepo, docs, CI, and a spike: a minimal runner with the Anthropic adapter and Electron streaming two conversations in parallel with cancel.
Done when: The spike works and the architecture is confirmed in an ADR.
Roadmap
From one desktop to a whole team.
Comitiva is built in phases. Each phase ends when a test exercises the whole path, not when the code is written.
- Phase 1Done
API connections
Anthropic, OpenAI-compatible, Google Gemini and Ollama; secrets in the OS keychain; the connections screen.
Done when: Four providers test and list models.
- Phase 2Done
CLI harnesses
Claude Code and Codex with their own login, streaming and session resume.
Done when: A turn with streaming and cancellation in both.
- Phase 3Done
Agents
Create and edit agents: role, connection, model, avatar.
Done when: The agent shows up in the sidebar.
- Phase 4Done
Chat
Full chat with parallel runs, persistence, retry and automatic titles.
Done when: Two agents respond at the same time.
- Phase 5Done
Tools
MCP in the runner, the tool loop for APIs, the filesystem server with roots and approvals, and tools for CLI harnesses through the runner.
Done when: An agent reads and creates a file in an allowed folder, a write asks for approval, outside the folder is denied.
- Phase 5bDone
Google Drive and MCP servers
OAuth, a built-in Google Drive server, and registering any MCP server in the app.
Done when: An agent reads a Google Doc and creates another one with approval.
- Phase 6Done
Usage
Usage records, a versioned pricing table, a dashboard and export.
Done when: The dashboard matches the records.
- Phase 7Done
Polish and v0.1.0
Attachments, search, export and import, i18n, installers for three systems, auto-update.
Done when: The release is published.
- Phase 8Done
Hub
A Laravel hub in its own open source repository (Apache-2.0, self-hostable): accounts, workspaces, shared agents, live sync of conversations, and a remote backend in the desktop. Each turn still runs on a member’s computer, with their own connection.
Done when: Two desktops see the same conversation live.
- Phase 8bDone
Guidelines
Blocks of instructions, rules or reference data for every agent, the agents with a tag, or the agents that pick them, added to the system prompt after the role. Import and export as AGENTS.md. In v0.3.0.
Done when: An agent’s run gets its role and then exactly the guidelines that reach it, in order, after importing them from an AGENTS.md.
- 9Phase 9Next
Web
The same UI in the browser, served by the hub. API connections and HTTP MCP servers run in the hub with team keys; a desktop can act as the workspace runner. Workspace guidelines, the same for every member.
Done when: A user without the desktop talks to a team API agent.
- 10Phase 10Planned
Usage policies
Limits, concurrency, fallback and connection switching.
Done when: An agent switches connection when it hits a limit.
- 11Phase 11Planned
Agents calling agents
Mention @Agent, or let an agent decide, and it asks another agent for help. The called agent answers in a linked conversation, with its own tools, approvals and usage; each agent lists the agents it may call.
Done when: An agent asks another agent and uses its answer, and the called agent’s approval shows up in the caller’s conversation.
- 12Phase 12Planned
Agent contacts
Share an agent as a link. The hub serves it as a remote MCP server with an A2A Agent Card, so another person’s AI — Comitiva, Claude, ChatGPT — can talk to it. A token per contact, revocation, usage limits, no write tools by default, and the owner sees the conversations.
Done when: Another person’s Comitiva adds the link as a tool server, and its agent gets an answer from the shared agent run by the hub.