Agentrove: Self-hosted AI workspace for orchestrating coding agents Agentrove, a self-hosted AI coding workspace from Mng-dev-ai, orchestrates Claude Code, Codex, Copilot, Cursor, Grok, and OpenCode agents from one interface, with Docker, macOS, and iOS apps. It uses an MCP server to let chats spawn sub-threads, assign per-turn models and personas, and run parallel worktrees for multi-agent workflows. The tool requires Docker and Docker Compose, and is available via GitHub at https://github.com/Mng-dev-ai/agentrove.git. Self-hosted AI coding workspace for running and orchestrating Claude Code, Codex, Copilot, Cursor, Grok, and OpenCode agents from one interface. - Runs Claude, Codex, Copilot, Cursor, Grok, and OpenCode through ACP adapters. - Gives each workspace its own Docker or host sandbox. - Combines chat, code editor, terminal, file tree, diffs, secrets, and git tools in one workspace. - Supports workspaces from empty folders, git clones, existing local folders, or GitHub repositories. - Streams agent sessions with cancellation, permission prompts, queued follow-up messages, file mentions, slash commands, and attachments. - Includes sub-threads, pinned chats, worktree mode, personas, custom instructions, environment variables, and installed agent skills. - Orchestrates multi-agent workflows through the bundled MCP server: a lead chat spawns worker sub-threads on any installed agent, model, and persona — in parallel worktrees when needed — then reviews their results. - Provides GitHub-assisted repository browsing, pull request review, PR creation, reviewer selection, and git branch/commit/push/pull helpers. - Ships as a Docker web app, a macOS desktop app, and a native iOS app. Agentrove chats aren't just endpoints — they can drive each other. The bundled MCP server mcp-server/ exposes the whole instance as tools send message , get messages , list models , list personas , … , and every chat's agent has those tools available. That turns any chat into an orchestrator: a lead agent on a strong model that decomposes the work, routes each task to the right agent, and reviews what comes back. The primitives: Sub-threads — send message parent chat id=… creates a worker chat grouped under the lead chat, in the same workspace and branch. Sub-threads stay flat no nesting : the lead fans out, workers report back. Per-turn model and persona — each worker runs on any installed agent and model model id with any persona a custom system prompt . Typical fleet: a fast model with a read-only scout persona for codebase exploration, coding models for implementation, dedicated reviewer personas bug hunting, structural quality for QA. list models reports each model's supported reasoning tiers thinking modes , so the lead dials effort per task. Isolated worktrees — worktree=true gives a worker its own git worktree, so parallel workers edit concurrently without conflicts. Polling and follow-ups — the lead polls get messages until a worker's turn completes, judges the result against the actual code, and sends rework to the worker's own thread; follow-ups inherit the thread's previous model, persona, and reasoning settings. Unattended turns — orchestrated turns run in the agent's full-execution mode, so workers finish without permission prompts. A typical loop: the lead explores with a cheap fast model, hands a precise spec to a coding model in a sub-thread, fans reviewer personas out over the diff in parallel, triages their findings, delegates the accepted fixes — and owns the final result. Models, personas, and routing rules are all user-defined, so the same machinery drives whatever fleet you run. Requirements: - Docker - Docker Compose git clone https://github.com/Mng-dev-ai/agentrove.git cd agentrove cp .env.example .env Set SECRET KEY in .env : openssl rand -hex 32 Start Agentrove: docker compose up -d Open http://localhost:3000 http://localhost:3000 . Agentrove also has a macOS desktop app built with Tauri. It starts a bundled Python backend sidecar on an available 127.0.0.1 port and connects the frontend to it at launch. - Download the latest Apple Silicon build from Releases https://github.com/Mng-dev-ai/agentrove/releases/latest . - Build from source: cd frontend npm install npm run desktop:dev Agentrove also builds a native iOS app with Tauri. Since iOS can't run the local backend sidecar, the app is a thin client: it talks to an Agentrove instance you already host your Docker or production deployment , reachable from the phone over https / wss . Because the project is open source, you build and sign it yourself — nothing is hardcoded to anyone else's server. Requirements: macOS with Xcode plus its iOS SDK and Simulator , the Rust iOS targets rustup target add aarch64-apple-ios aarch64-apple-ios-sim , and CocoaPods. If your device runs an iOS beta whose SDK isn't in a stable Xcode yet, build against the latest stable Xcode's SDK sudo xcode-select -s /Applications/Xcode.app — apps built with an older SDK still run on newer iOS. Point the app at your instance: cd frontend cp .env.mobile.example .env.mobile then set your https/wss URLs npm install Run in the simulator: npm run ios:dev Install on your own iPhone with a free Apple ID — no paid developer account needed the app must be re-signed every 7 days : npm run tauri ios init generates the Xcode project first run only .- Open frontend/src-tauri/gen/apple/ .xcodeproj in Xcode once, pick your Team under Signing & Capabilities , and connect your iPhone Xcode needs to register the device and create the provisioning profile . - On the phone, enable Developer Mode Settings → Privacy & Security , then after the first install trust the certificate under Settings → General → VPN & Device Management . Build a standalone .ipa and install it on the connected iPhone. The easiest path is the bundled helper, which builds, signs, exports, and installs in one step. It reads your team from APPLE DEVELOPMENT TEAM so nothing is hardcoded to anyone else's team and auto-detects the connected device: export APPLE DEVELOPMENT TEAM=