Multica
TL;DR
Multica is a source-available task board from Index Labs where people assign issues to coding agents such as Claude Code, Codex or OpenCode, which run on machines you connect and report back for review. Its Multica License adds hosted-service and branding conditions to Apache-2.0. It suits teams that want ticket-driven agent work, self-hosted or cloud.
Key facts
| Type | Platform |
|---|---|
| Languages / SDKs | Go, TypeScript |
| License | unknown |
| Pricing model | Open core |
| Orchestration pattern | Supervisor |
| GitHub stars | 51,737 (as of 2026-09-30) |
| GitHub forks | 6,718 |
| Last push | 2026-09-30 |
| Latest release | v0.6.0 |
| Repository | multica-ai/multica |
| Website | multica.ai |
| Documentation | multica.ai |
| Last verified | 2026-09-30 |
Key features
- Agents appear as assignees on a board: assigning an issue, @-mentioning an agent, chatting with it, or an Autopilot schedule each create a run. (source)
- A daemon on your own computer or cloud box claims runs and starts the local agent CLI; the README lists 26 supported CLIs, including Claude Code, Codex, OpenCode, OpenClaw and Hermes. (source)
- Squads put a leader agent in charge of routing: the leader reads the issue, delegates to members with an @-mention, and is re-triggered when a member reports back. (source)
- Every run gets its own working directory, and the run's API token is bound to that agent and run. (source)
- Issues have built-in statuses including in_review and blocked; the README says work lands in review, not in main. (source)
- Execution logs replay every tool call and error, and a reply to a running agent lands in the current run for Claude Code, Codex and Grok. (source)
- MCP server settings can be stored on an agent and are passed to tools that support managed MCP before a run. (source)
- The
multicaCLI and an API script issues, agents and skills; agents drive Multica through the same CLI. (source)
Architecture and orchestration pattern
Pattern: Supervisor
Multica has a server that stores workspaces, issues, comments, agent configuration, skills and run records, plus a daemon that runs on each connected computer. Assigning an issue creates a run in a queue; an online runtime claims it, launches the configured agent CLI in a per-run working directory, streams progress back and writes the result to the issue. A run is bound to its runtime and never moves to another machine.
The agent tools are the user's own installed CLIs; Multica does not ship a model. It records and schedules, while file changes and commands execute on the connected computer. Custom environment variables and MCP configuration saved on an agent are the exception: they are stored on the server and sent to the runtime.
Multi-agent work is issue-driven. A squad has a leader agent and members (agents or people). Assigning an issue to a squad wakes only the leader, which posts a delegation comment mentioning members; each mention triggers a new run. Agents never start on their own: every run comes from an assignment, mention, chat message or Autopilot.
Human in the loop
Humans are teammates on the same board. Issues have an in_review status and the README says work lands in review rather than in main, so a person decides what ships. A user can steer a running agent by replying while it works (Claude Code, Codex and Grok), cancel a run, and read the execution log. Agents raise blockers and the Inbox pings members when a decision is needed. Approval prompts of the underlying tool are answered automatically because runs are unattended.
Harnesses it can drive
- Claude Code (evidence)
- Codex (evidence)
- OpenCode (evidence)
- OpenClaw (evidence)
- Hermes (evidence)
Protocols
| Protocol | Support | Note |
|---|---|---|
| MCP | Partial evidence | The docs say MCP servers can be defined on an agent and Multica passes them to the tool before a run, but only for tools with managed MCP (Antigravity, GitHub Copilot CLI, DevEco Code and Pi are excluded). This is configuration pass-through; no Multica-hosted MCP server or MCP client role is described. |
| A2A | Unknown | Searched the README, docs pages under apps/docs/content/docs and the repository file list for 'a2a' and 'agent2agent'; nothing found. Agent-to-agent handoff uses issue comments and squads. |
| AG-UI | Unknown | Searched the README, docs and file list for 'ag-ui'; nothing found. |
Best for
- Teams that already work from an issue board and want Claude Code, Codex or OpenCode to pick up assigned issues and return them for review. (shortlist)
- Running many agents across several machines or a cloud box you control, with run logs and usage analytics per issue. (shortlist)
- Self-hosting the whole stack on Docker Compose or Helm, with roles and access scopes for members and agents. (shortlist)
- Recurring agent jobs such as audits and standup reports through Autopilots. (shortlist)
Not for
- Offering Multica to outside users as a hosted service or embedding it in a product you sell: the license requires a commercial license.
- Users needing a sandbox by default: the docs say runs execute with the daemon user's full permissions and Multica gives no filesystem sandbox.
- Solo developers who want a local terminal or IDE workflow rather than a task board.
Quickstart
curl -fsSL https://raw.githubusercontent.com/multica-ai/multica/main/scripts/install.sh | bash -s -- --with-server
multica setup self-host# Self-host the server and connect this machine as a runtime (requires Docker)
curl -fsSL https://raw.githubusercontent.com/multica-ai/multica/main/scripts/install.sh | bash -s -- --with-server
multica setup self-host
multica auth status
multica daemon status
# Agents are created in the web UI (Agents > New agent); then drive issues from the CLI
multica issue create --title "Fix login failure"
multica issue assign MUL-123 --to "Backend Agent"
multica issue runs MUL-123
multica issue comment add MUL-123 --content "Check the regression tests first"
Common pitfalls
- The machine that runs agents needs at least one supported agent CLI installed and signed in; Multica drives them but does not ship them.
- The self-host install pulls images from GHCR and requires Docker; if the selected tag is not published yet, fall back to
make selfhost-buildfrom a checkout. - Runs execute with the full permissions of the daemon's OS user, and approval prompts are answered automatically; run the daemon as a dedicated user, in a container or in a VM.
- Anything saved to an agent's custom environment variables or MCP configuration is stored on the Multica server, not only on your machine.
- A run is bound to its runtime; if that runtime is offline the run waits in the queue.
- Windows uses the PowerShell installer; open Windows issues report runner and login problems.
- Tools without managed MCP (Antigravity, GitHub Copilot CLI, DevEco Code, Pi) ignore MCP settings saved on an agent.
Pros
- Uses the agent CLIs and accounts you already have; the docs say Multica does not ship a model and the README lists 26 supported CLIs. (source)
- Board-native workflow with run queue, retries, execution logs and per-run working directories, so concurrent agents do not share a checkout. (source)
- Squads give a leader agent the job of routing work to members, including human members. (source)
- Self-hostable with Docker Compose or Helm, or usable as a hosted cloud service; the source is public. (source)
- Frequent releases (v0.6.0 on 2026-09-28) and clear security documentation of what a run can reach. (source)
Cons
- Source-available, not open source: the Multica License adds a hosted-service ban, branding conditions and a contributor clause on top of Apache-2.0. (source)
- Runs use the daemon user's full permissions with no filesystem sandbox, and unattended runs answer approval prompts automatically. (source)
- Configuration such as custom environment variables and MCP settings is stored server-side, so not everything stays on your machine. (source)
- Squads route work but do not raise concurrency, and agents never start on their own; every run needs an explicit trigger. (source)
- Still 0.x, and the iOS app builds from source only; some chat channels are community-maintained. (source)
Alternatives
FAQ
Does Multica support MCP?
Partly. You can store MCP server configuration on an agent and Multica passes it to tools that support managed MCP before a run. It is not described as an MCP server, and a few tools ignore the setting.
What license is it under?
GitHub reports no recognized license. The LICENSE file is the Multica License: the full Apache License 2.0 text plus extra conditions. Internal use, including self-hosting for your own organization, is allowed; offering it to third parties as a hosted service or embedding it in a product you sell needs a commercial license, and the UI's logo and attribution must be kept unless waived.
Is Multica free?
Self-hosting for your own organization is free under the license, and there is a hosted cloud version. The homepage shows a free trial and a sales contact; no pricing page was found, so plan prices are not stated here.
Which coding agents can it run?
The README lists Claude Code, Codex, OpenCode, OpenClaw, Hermes, Cursor Agent, Copilot CLI and many more, 26 in total. Each must be installed and signed in on the runtime machine.
How does it differ from Symphony?
Symphony polls an existing tracker and runs Codex per issue as a background service. Multica is its own board with agents as assignees, squads, chat, run logs and a runtime daemon, and it supports many agent CLIs.