AgentDock
kouyusanpi/coding-agent-dock
Orquestrador desktop para múltiplos agentes de código IA com cluster coordination
Ver no GitHubAplicação desktop Flutter que gerencia e coordena múltiplos agentes de IA (Claude Code, Gemini, Copilot, etc.) em um único workspace unificado. Executa sessões concorrentes em terminais PTY embutidos, sincroniza memória compartilhada entre agentes e fornece camada de orquestração via API IPC para handoff de tarefas.
Instalação
npx convez-ai kouyusanpi/coding-agent-dock1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Executar múltiplos agentes de IA simultaneamente em um projeto
- Manter contexto compartilhado entre diferentes ferramentas de código (Claude, Gemini, Copilot)
- Automatizar handoff de tarefas entre agentes via IPC
- Gerenciar sessões Claude com resumo automático e persistência
Veredito

Ferramenta sofisticada para quem usa múltiplos agentes de IA em fluxos complexos. A orquestração via IPC e sincronização de memória compartilhada é inovadora, mas exige macOS/Linux/Windows e CLI agents já instalados — não é para iniciantes.
README
A tradução para o português deste README ainda está na fila.
AgentDock
A Flutter desktop app for managing, running, and coordinating multiple local AI coding agent CLIs in a single unified workspace. Supports macOS (primary), Linux, and Windows.
Bundle:
AgentDock.app· Identifier:com.plugins.coding_agent_dock· Flutter 3.x · macOS 10.15+ / Linux / Windows 10+
Overview
AgentDock detects every AI coding agent CLI installed on your machine, lets you run concurrent task sessions in embedded PTY terminals, and provides a cluster coordination layer so agents can communicate, share state, and hand off work to each other — all without leaving the app.
Supported Agents
| Agent | Provider | Install |
|---|---|---|
| Claude Code | Anthropic | npm i -g @anthropic-ai/claude-code |
| OpenAI Codex CLI | OpenAI | npm i -g @openai/codex |
| Gemini CLI | npm i -g @google/gemini-cli | |
| GitHub Copilot | GitHub | brew install gh && gh extension install github/gh-copilot |
| Aider | Aider-AI | pip install aider-install && aider-install |
| Amazon Q Developer | AWS | brew install amazon-q |
| Goose | Block | install script from GitHub |
| Plandex | Plandex AI | install script from plandex.ai |
| Amp | Sourcegraph | npm i -g @sourcegraph/amp |
| Open Interpreter | — | pip install open-interpreter |
| GPT Engineer | — | pip install gpt-engineer |
| Cursor CLI | Cursor | enable from Cursor app |
| Windsurf | Codeium | windsurf.com |
| Continue CLI | Continue | npm i -g @continuedev/cli |
| CodeWhale | — | cargo install codewhale |
Custom CLIs can be added in the app — any binary on your PATH works.
Features
Terminal Management
- Concurrent sessions — run multiple agents simultaneously in independent PTY terminals with xterm rendering
- Claude session resume — every Claude session gets a UUID; the app automatically passes
--session-idon first launch and--resumeon reopens, preserving conversation context across restarts - Live status dots — per-session indicators (running / completed / failed / cancelled) update in real time
- Session pinning — pin frequently-used sessions to the top of the list
- Sort & filter — sort by name, date, or status; filter by agent type or running state
Project Grouping & Shared Memory
- Sessions are grouped by working directory, giving each project its own visual cluster
- Shared memory file — create
.agentdock/shared-memory.mdin a project directory; AgentDock syncs its contents into each agent's native memory file (CLAUDE.md,AGENTS.md,GEMINI.md) inside a managed block, so every agent in the same project shares the same context - Live sync polls running sessions every 5 s and re-injects if the file changed
- Status dot — a green indicator on the project group header shows at a glance whether shared memory is active for that project
- Shared Memory editor — click the brain icon on any project group to view or edit the shared memory file directly from the app
Cluster IPC API
Every spawned PTY receives three environment variables:
AGENTDOCK_API_BASE = http://127.0.0.1:<port>/v1
AGENTDOCK_IPC_URL = http://127.0.0.1:<port>/v1/sessions/<id>/events
AGENTDOCK_SESSION_ID = <sessionId>
AGENTDOCK_HELPERS = ~/.agentdock/helpers.sh
Agents can source "$AGENTDOCK_HELPERS" in a Claude Code hook or shell script to get a set of ready-made bash functions:
| Function | Description |
|---|---|
agentdock_list | List all sessions (JSON) |
agentdock_status SESSION_ID | Get a single session's status |
agentdock_wait SESSION_ID [TIMEOUT] [POLL] | Block until a session finishes |
agentdock_output SESSION_ID [MAX_LINES] | Read the last N lines of a session's output |
agentdock_stream SESSION_ID | Stream live output (SSE) |
agentdock_inject SESSION_ID "message" | Write text into another agent's stdin |
agentdock_notify [TYPE] [DATA_JSON] | Post an event to this session's endpoint |
agentdock_running_ids | Print IDs of all running sessions |
agentdock_broadcast "message" | Send a message to all other running sessions |
agentdock_kv_get KEY | Read a value from the shared KV store |
agentdock_kv_set KEY VALUE [TTL_SECONDS] | Write a value (optional auto-expiry TTL) |
agentdock_kv_del KEY | Delete a key |
agentdock_kv_list | List all live keys |
The shared KV store (GET/POST/DELETE /v1/kv/:key) is an in-memory store that persists for the lifetime of AgentDock. Use it to pass structured state between agents without touching the filesystem — values can carry an optional TTL so they auto-expire.
Auto-Relay Pipeline
Chain sessions together: hover a running session and set a "chain-to" agent. When the session exits with code 0, AgentDock immediately opens a new session for the chained agent in the same working directory — zero-click handoff between Claude Code → Codex → any other agent.
Cluster Comparison
Run the same prompt on multiple agents at once ("Run on all") and compare their outputs side by side in the Cluster Comparison Dialog. Sessions in the same batch share a batchId so the comparison panel always groups them correctly.
Arg-Error Self-Healing
If a session exits with an "unknown option" error (common when CLI flags change across versions), AgentDock:
- Parses the rejected flag from the error output
- Strips just that flag from future launches and persists it per CLI in settings
- Automatically relaunches — no user interaction needed
Denied flags are cleared automatically when a newer version of the CLI is detected.
Custom Launch Arguments
The new-session dialog has a Custom launch arguments toggle. Enter raw CLI flags (quote-aware tokenization) or leave blank for a bare command with no flags. Override or extend the auto-generated flag set for any session.
Event Log
A chronological ring-buffer (last 200 events) tracks every significant cluster event: session starts/stops, IPC notifications, pipeline relays, watchdog retries, memory syncs. Open it with ⇧⌘L or from the Settings drawer.
CLI Detection Diagnostics
When a CLI is not found, the install dialog shows every path that was searched (common paths + full PATH scan) so you can see exactly why detection failed. A one-click copy button on the install hint makes it easy to run the correct install command.
Build & Run
# Run in development
flutter run -d macos
# Static analysis
flutter analyze
# All tests (321 tests)
flutter test
# Regenerate Drift ORM code (after editing tables.dart or database.dart)
dart run build_runner build --delete-conflicting-outputs
# Regenerate localizations (after editing lib/l10n/*.arb)
flutter gen-l10n
# Package a release build → dist/AgentDock-<version>.{app,zip,dmg}
scripts/package_app.sh
macOS: Universal Binary Note
The macOS build ships as a universal binary (arm64 + x86_64, macOS 10.15+). pubspec.yaml pins path_provider_foundation: 2.4.4 — do not bump past 2.4.x until Flutter ships multi-arch native-assets for macOS, or Intel builds will break. scripts/package_app.sh enforces this by failing if any bundled framework is missing x86_64.
macOS: App Sandbox
The App Sandbox is deliberately disabled in both entitlements files — the app must exec local CLI binaries and spawn PTYs. Do not re-enable it.
Linux / Windows
Both platforms include platform directories and Flutter's PTY support works on all three desktops. Primary development and QA is on macOS; Linux and Windows builds are community-supported.
Architecture
CliRegistry static list of known CLIs + detection metadata
↓
CliDetector locate binary (which → commonPaths → PATH scan)
↓
CliCacheService JSON file cache of detection results
↓
HomeScreen single-page workspace (sidebar + terminal pane)
↓
NewSessionDialog per-session config (agent, prompt, working dir, custom args)
↓
SessionManager creates DB record, builds CLI flags (Claude settings → flags,
custom args, denied flags stripped)
↓
TerminalSessionsController owns all concurrent PTYs (flutter_pty + xterm)
↓
AppDatabase (Drift/SQLite) persists sessions, status, custom args
Storage tiers — do not mix:
| Tier | Service | Contents |
|---|---|---|
| SharedPreferences | SettingsService, ClaudeSettingsService | UI prefs, per-CLI settings, denied flags |
| JSON file | CliCacheService | CLI detection results (non-critical) |
| SQLite (Drift) | AppDatabase | Task sessions, status, history |
IPC flow: IpcServer binds to a random localhost port on startup; all PTYs receive the port as AGENTDOCK_API_BASE. Events posted by agents arrive as a broadcast stream consumed by HomeScreen._onIpcEvent.
Directory Structure
lib/
├── database/ Drift schema + migrations (schema v8)
├── l10n/ Localizations (en + zh)
├── models/ AgentCli, CliRegistry, TaskSession
├── screens/ HomeScreen, NewSessionDialog
├── services/ Detection, sessions, IPC, KV, memory sync, workflows
├── theme/ AppColors, AppTypography, AppSpacing
├── utils/ ANSI stripping, shell tokenizer
└── widgets/ TaskPanel, TerminalPane, dialogs
Localization
UI strings are in lib/l10n/app_en.arb (English) and lib/l10n/app_zh.arb (Chinese). Add new strings to both files, then run flutter gen-l10n.