mrkeyoor.com_
Fri 25 Sept 21:45 UTC
Automationevaluationupdated 27 Aug 2026

multica review

Multica is a work board for assigning software tasks to AI coding agents and keeping their runs, comments, blockers, costs, and human review in one issue. It drives 23 supported agent command-line tools on machines you control, while the board can run in Multica's cloud or on your own server.

+744stars / 7d
Verdict

Our Multica run installed 2,014 packages, used 2,385 MB, and spent 539 seconds building and testing successfully, so this is serious team infrastructure rather than a lightweight agent launcher. Adopt it when multiple coding CLIs already create a coordination problem and a human review gate is non-negotiable. Wait if you need fine-grained project privacy, encrypted agent environment values, or broad commercial hosting rights under a standard open-source license.

We ran it

Lab card: what happened when we ran multicaScreenshot of multica (multica.ai)
Install✓ · 88s2014 packages · 2385 MB
Build✓ · 180s
Tests✓ · 359sran, no count parsed
Repo5134 files~1,031,677 lines of source · 50.7 MB · 4 CI workflows · Dockerfile

Answers from our run

Does multica build from source?

Dependencies installed in 88 seconds (2014 packages), and the build succeeded in 180 seconds. We cloned commit 1cc46b2 into a clean Debian container with 3 CPUs and no project-specific setup.

Do multica's tests pass?

The test command failed in our container, and its output did not report a pass or fail count.

Who should not use multica?

Companies planning to offer Multica to outside users without a commercial agreement: its added license conditions restrict hosted and embedded third-party services.

What are the alternatives to multica?

OpenHands, Dify, Langflow. Our Multica run installed 2,014 packages, used 2,385 MB, and spent 539 seconds building and testing successfully, so this is serious team infrastructure rather than a lightweight agent launcher.

Setup3/5Build passes; server, daemon, database, and agent CLIs remain
Docs5/5Cloud, Docker, Helm, daemon, auth, and security paths are detailed
Community5/547,880 stars and daily code, issue, and release activity
Maturity3/5Fast releases coexist with current runtime and permission defects

Discussed on

  1. hnYour next 10 hires won't be human3 points

Who it’s for

Engineering teams already using several coding agents and losing track of who ran what.
Technical leads who want agent work to end in a review queue instead of landing directly in the main branch.
Self-hosters with Docker or Kubernetes skills who need agents to execute beside private source code.
Teams willing to manage provider accounts, agent CLI authentication, runtime access, and task costs.

Who it’s NOT for

Companies planning to offer Multica to outside users without a commercial agreement: its added license conditions restrict hosted and embedded third-party services.
Teams that need project-level member isolation today: issue 6064 says workspace members can see projects and other people's issues.
Security programs that prohibit plaintext application secrets in a database: issue 7595 says agent custom environment values are stored as plaintext JSONB.
Operators expecting every listed runtime to behave alike across operating systems: current reports cover OpenClaw preparation failure, Windows token injection, and Windows path overrides.
Small teams using one agent in one terminal: the board, daemon, PostgreSQL, and review workflow add more machinery than that job needs.

Setup reality

Our sandbox installed 2,014 pnpm packages in 88 seconds and used 2,385 MB. The build succeeded in 180 seconds, and tests succeeded in 359 seconds. Commit 1cc46b2 had 5,134 files and about 1,031,677 source lines.

Self-hosting needs Docker Compose or Kubernetes, PostgreSQL 17, a JWT secret, and a CLI daemon on every machine that runs agents. Email delivery uses Resend, though a local trial can retrieve login codes from backend logs. Each agent CLI needs its own installation and authentication.

Contributors need Node.js 22, pnpm 10.28.2, Go 1.26.6, and Docker. The daemon executes beside the code and talks to the server over WebSockets. Public or embedded third-party service use is restricted by additional license terms.

Twenty-three agent CLIs share one issue board

Multica treats Claude Code, Codex, Cursor, Copilot, Kimi, OpenCode, and 17 other command-line agents as workers on the same board. Assign an issue and the chosen agent claims it, runs on a registered machine, comments on progress, and returns the issue for review. The execution log keeps commands, errors, timestamps, and token use with the task. That solves a specific management problem: terminal sessions stop being the only record of what an agent attempted.

The board also includes squads, scheduled autopilots, chat, projects, inbox notifications, roles, and reusable skills. Release v0.4.35 shipped on August 26, 2026, with fixes for Codex turn completion, runtime access, private agents, and autopilot assignment. Multica is broad enough to become an operating layer for agent work. A developer using one agent interactively will spend more time configuring the layer than recovering lost context.

The daemon keeps code local while the server keeps control state

A Go backend and Next.js 16 frontend store workspace state in PostgreSQL 17. The separate daemon runs on a developer laptop or cloud machine beside the repository and launches the chosen agent CLI. Multica does not include the models or agent subscriptions. Every runtime must already be installed, authenticated, and able to reach its provider. Code execution stays on the runtime you select, while task state and logs go to the Multica server.

That split is useful for private repositories, but it creates 2 security boundaries. Server roles decide who may invoke a runtime, and the local daemon holds the practical ability to run commands near source code. Issue 7579 describes a gap when a public runtime becomes private: teammate agents can remain bound even though a claim guard blocks new work. Review runtime visibility, queued tasks, environment injection, and cancellation behavior with the exact agent CLIs your team permits.

What happened when we ran it

Our sandbox installed 2,014 pnpm packages in 88 seconds and consumed 2,385 MB on disk. commit 1cc46b2 contained 5,134 files, roughly 1,031,677 source lines, and a 50.7 MB checkout. The build succeeded in 180 seconds. This is a large monorepo spanning web, desktop, mobile, backend, daemon, shared packages, and documentation, so the dependency size matches the product's surface even if it remains costly for contributors.

Tests succeeded in 359 seconds in the same unprivileged container with 3 CPUs, 8 GB of RAM, Node 22, and no secrets. The repository had 4 CI workflow files, a Dockerfile, a compose file, and workspace configuration, though our scan found no top-level tests directory. The measured success covers the commands run at commit 1cc46b2. It does not prove a provider login, a live daemon connection, or successful task execution through any of the 23 adapters.

Docker starts three services, then every runner needs a daemon

The self-host quick path installs the CLI, pulls server images, creates configuration, and starts the stack. Underneath, the deployment has a Go API with WebSockets, a Next.js frontend, and PostgreSQL 17. Each person who wants local agent execution then installs the multica CLI, at least 1 supported agent CLI, and a background daemon. Login can use Resend email codes; a private trial may read generated codes from backend logs.

Kubernetes users get a Helm chart with 2 ingress routes, a PostgreSQL volume, and a separate uploads volume unless S3 is configured. The chart expects operators to create the secrets object themselves. Contributors face another stack: Node.js 22, pnpm 10.28.2, Go 1.26.6, and Docker. make dev automates much of the setup, but the successful 539 seconds of build and tests in our run should set expectations for change feedback.

Current permission and secret gaps matter for shared teams

Issue 6064 says members can see projects and issues outside their assignments and asks for project and issue visibility controls. The README documents owner, admin, and member roles plus runtime access scopes, so this is a narrower gap than having no authorization. It is still important for agencies, internal divisions, or any workspace where issue text contains customer information. Verify visibility with two ordinary member accounts before importing sensitive work.

Issue 7595 says custom agent environment values are stored as plaintext JSONB, although API reads redact them and agent tokens cannot retrieve them. Database backups and read-only database access therefore deserve the same protection as live provider credentials. Other August reports are adapter-specific: Windows token injection blocks DeepSeek Harness, an executable check rejects Windows path overrides, and OpenClaw configuration preparation can fail before a task reaches the model.

The license allows internal use but restricts outside hosting

GitHub could not map the repository to a standard SPDX license because Multica reproduces Apache 2.0 with added conditions. Internal use within one organization is allowed. Without a commercial license, the terms prohibit using the source to provide a hosted service to third parties or embedding it in a commercially distributed product. The interface branding also cannot be removed without a separate waiver, and non-interface uses retain attribution duties. Legal review is appropriate before customer-facing deployment.

The project itself is moving quickly. GitHub showed 47,880 stars, 1,331 combined open issues and pull requests, and a last push on August 27, 2026. The v0.4.35 release arrived the previous day, and the README says releases happen most weekdays. That activity improves the chance of fixes while increasing upgrade churn. Multica earns a pilot for a team already coordinating several agents, provided the pilot tests access, secrets, cancellation, and every chosen runtime on its real operating system.

Alternatives

ProjectWhat it isPick it when
OpenHands gh↗A software development agent platform with a web interface and sandboxed execution.pick this instead when you want one primary agent platform rather than a board that coordinates many existing CLIs.
Dify gh↗A visual platform for building and operating general AI applications and workflows.pick this instead when business workflows and deployed AI apps matter more than issue-driven coding agents.
Langflow gh↗A visual builder for agent and retrieval flows with reusable components.pick this instead when you need to design model flows rather than supervise coding work across repositories.

What people are saying

  1. [github-trending] multica-ai/andrej-karpathy-skills
  2. [github-trending] multica-ai/multica

Sources

  1. Multica repository
  2. Multica self-hosting guide
  3. Multica license
  4. Multica v0.4.35 release
  5. Member visibility request
  6. Agent environment encryption request
  7. Runtime visibility bug

More automation reviews

Rose · alchemy · laya · prod-FARM-IOS-Core · ha_xiaomi_home · FreePEP · the whole board →