preset-io/agor

Agor – Multiplayer AI Workspace

What it is – Agor is a self‑hosted (or managed‑cloud) web platform that lets software teams work together with large‑language‑model coding agents (Claude Code, Codex, Gemini, Copilot, etc.). Each piece of work lives on its own Git branch, which gets an isolated dev environment, a persistent conversation history, and a live visual “board” where teammates and agents can see each other’s cursors, comments, and artefacts in real time.

Core concepts

• Branches – a Git branch plus a dedicated dev server and chat history; the basic unit of work.
• Sessions & Trees – agent‑to‑agent or agent‑to‑human conversations that can be forked or spawned, forming a genealogy of prompts.
• Boards & Zones – a 2‑D canvas where branch cards are placed; dropping a card into a zone can fire a templated prompt.
• AI teammates – persistent agents with their own searchable knowledge‑base, skills, and scheduled tasks. They can be reached via Slack, GitHub, or directly on the board.
• MCP server – an HTTP endpoint (/mcp) that lets agents introspect and control Agor (fork, spawn, schedule, report) using a simple JSON‑RPC protocol.

Key capabilities

Feature	What you get
Multiplayer & social	Live cursors, face‑piles, spatial comments, shared terminal so everyone can follow an agent’s work.
Boards & Zones	Visual organization of branches; zones trigger reusable prompts when a branch is dropped.
Persistent teammates	Durable memory (markdown KB), skills, and tool bindings; agents can act autonomously.
Knowledge base	Shared searchable markdown store for decisions, runbooks, prompts, and agent memory.
Message gateway	Interact with teammates from Slack or GitHub without opening the UI.
Scheduler	Cron‑style triggers for recurring prompts (stand‑ups, audits, digests).
Rich chat UX	Token and dollar accounting per prompt, model/effort selectors, structured tool output blocks.
One‑click dev environments	Automatic dev server per branch with unique ports, so any teammate can inspect the running build.
Security & RBAC	Branch‑scoped permission tiers, per‑user credentials/env vars, and three execution modes (simple, sandbox, delegated).
Artifacts	Live interactive apps (dashboards, mockups) rendered directly on the board.

How it works – The system consists of a daemon (FeathersJS) that owns a SQLite/Postgres database, Git worktrees, and the MCP HTTP endpoint. A separate executor process runs the selected agent SDKs (Claude Code, Codex, Gemini, OpenCode, Copilot, Cursor) in isolated sandboxes. The UI (React + Ant Design) and CLI (oclif) talk to the daemon via REST/WebSocket.

Installation (quick‑start)

# Prereqs: Node ≥22.12 and Git
npm install -g agor-live
agor init            # pick which agent SDKs to install, creates config & DB
agor daemon start    # runs the backend daemon
agor open             # launches the web UI in your browser

A Docker‑Compose variant is also provided for those who prefer containerised setup.

Typical use cases

• Team‑wide AI coding – Let every developer see what an LLM is doing on a feature branch, jump in to adjust prompts, and reuse successful workflows.
• AI‑assistant onboarding – Create a persistent teammate that knows your internal runbooks, can answer questions, and can run scheduled audits.
• Cross‑tool orchestration – Switch between Claude Code, Codex, Gemini, etc., per session without changing the surrounding infrastructure.
• Governance & cost tracking – Per‑prompt token and dollar accounting, RBAC per branch, and audit logs for compliance.
• Education & knowledge sharing – New hires can watch live agent sessions, see the prompts that work, and copy them into the shared knowledge base.

License – Source‑available under the Business Source License 1.1 (commercial use allowed, redistribution of the agent‑orchestration functionality restricted until a future Apache‑2.0 conversion date).

Community – Discord, GitHub Discussions, and Issues are the main support channels.

---

TL;DR – Agor is a collaborative platform that turns LLM coding agents into first‑class teammates, giving software teams a shared visual board, persistent memory, scheduling, and fine‑grained governance while remaining agnostic to the underlying LLM provider.