Skills Atlas · Multica Coding Agent issue-board collaboration platform (open source, supports 10 CLI agents, can be self-hosted)

Create issues with multica issue create and assign them to a colleague or agent. The agent picks up the issue, writes code, leaves blockers as comments, and opens sub-issues. PMs track human and agent progress on the same board without switching tools.

Agents stream progress through WebSocket throughout a task; the backend uses gorilla/websocket. You do not need to wait at the terminal. An agent proactively reports blockers when stuck and writes the result into the issue when the task succeeds.

multica daemon start starts the local runtime. The CLI automatically detects agents installed on PATH, including Claude Code, Codex, Cursor Agent, Gemini, Hermes, Kimi, Kiro, OpenCode, OpenClaw, and Pi, adding all of them to the routing list. Local and cloud runtimes can run simultaneously, with their status shown in one dashboard.

Each workspace is independently isolated, with issues, Skills, agent settings, and permissions completely separate. Agencies can create a separate workspace for each client; their Skills do not interfere with one another, while operations and accounts share one infrastructure. Suitable for freelance teams requiring client confidentiality.

curl -fsSL https://raw.githubusercontent.com/multica-ai/multica/main/scripts/install.sh | bash. The script detects the operating system and architecture, then places the binary on the appropriate PATH. Suitable for CI, Docker base images, and Linux servers without Homebrew.

irm https://raw.githubusercontent.com/multica-ai/multica/main/scripts/install.ps1 | iex. Windows developers do not need WSL or a third-party package manager first; PowerShell 7+ is sufficient. After installation, use multica setup for initial setup.

curl -fsSL .../install.sh | bash -s -- --with-server plus multica setup self-host. For manual deployment, use docker compose -f docker-compose.selfhost.yml up -d. The backend is a Go single-file binary using Chi router, sqlc, and gorilla/websocket; the database is PostgreSQL 17 + pgvector. The default ports are frontend 3000 / backend 8080. Generate JWT_SECRET with openssl rand -hex 32.

multica setup (initial setup, authentication, daemon startup) → multica daemon start (start the local runtime) → multica issue create (create an issue) → multica issue list (list all workspace issues). Together with the web dashboard, these handle everyday work without learning an additional SDK.

On its first run, multica setup scans agent CLIs installed on PATH, including Claude Code, Codex, Cursor Agent, Gemini, Hermes, Kimi, Kiro, OpenCode, OpenClaw, and Pi. Results appear directly in the dashboard, and every agent can be assigned an issue. After adding an agent, run multica setup again to detect it.

Generate JWT_SECRET with openssl rand -hex 32; do not use online generators or hardcoded strings. RESEND_API_KEY is the officially recommended email verification option, keeping self-hosted registration and login consistent with the cloud edition. For local testing, MULTICA_DEV_VERIFICATION_CODE can bypass email, but must never be enabled in production.

The backend uses Go + gorilla/websocket for live progress updates. During long tasks, the dashboard gradually displays actions, commits, and blockers. You can work on other tasks while the agent executes, then return to review its results after it commits. When stuck, the agent proactively writes the blocker into an issue comment.

Multica and forrestchang / andrej-karpathy-skills are two projects by the same author, Forrest Chang. Multica assigns issues to agents and manages execution; Karpathy CLAUDE.md defines agent behavior. Commit it into each Multica workspace's project repo, and agents automatically apply four principles during issue execution, reducing the chance of over-engineering and drive-by refactors.

Human and agent issues appear on the same board. During sprint planning, PMs assign automatable issues to agents and keep stakeholder-alignment tasks with humans. Agent comments, sub-issues, and blockers are visible at the same level as human activity; standard PM workflows need no changes.

Agents execute in a local daemon or self-hosted cloud, and code never passes through Multica servers. For scenarios involving NDAs, data-protection laws such as GDPR and the Personal Data Protection Act, or clients requiring code to remain internal, local execution is a prerequisite for compliant deployment. Apache 2.0 licensing makes procurement legal review easier to pass.

Multica CLI detection lets one workspace schedule Claude Code for complex refactoring, Codex for pure-function logic, Cursor Agent for small fixes inside the IDE, Gemini for large-context tasks, and Kimi or Hermes for regional-language or long-context scenarios. Specify a preferred agent when dispatching an issue and choose tools for the task type.

Tenten’s approach: create a workspace for each new client on one self-hosted Multica installation. Client A’s Shopify Plus, client B’s Webflow, and client C’s Headless Sanity knowledge accumulate separately without interference. There is one operations setup and one infrastructure; compared with deploying a separate tool for each client, costs and operational burden are clearly lower.

Multica reached GitHub Trending #5 during a certain period and gained 1,724 stars in one day. For a newly launched self-hosted tool, this growth primarily came from feature-driven sharing rather than advertising. In the same period, ARK Invest’s Big Ideas 2026 report said AI agent task duration grew 5–6 times during 2025, increasing demand for the platform layer.

AgentConn's long review positions Multica as a representative platform for bringing coding agents into issue-board collaboration. It focuses on the issue board and skill compounding, matching README's emphasis. The report says AI Coding Agents as Real Teammates describes the product difference more accurately than another agent runner.

Flowtivity writes from the perspective of turning AI into workflows, explaining the architectural significance of agents becoming collaboration units rather than individual tools and the effects on workflow automation and PMO toolchain integration. The article also emphasizes vendor neutrality and self-hosted deployment for enterprise procurement.

Arun Baby compares Multica and hermes-agent in his AI agents series. "The key difference is that hermes-agent is about an agent that grows with you, while multica is about a platform that helps your team grow with agents." This comparison positions Multica as the collaboration layer, rather than the agent itself. It does not compete directly with tools such as Claude Code and Cursor Agent; it is built on top of them.