@flightdeck-ai/flightdeck
Advanced tools
[!WARNING] This is purely AI generated code. Use the project with this understanding in mind.
One command. A whole engineering crew.
Flightdeck orchestrates crews of AI coding agents through a real-time web UI — each with a specialized role, its own context window, and the ability to collaborate through structured messaging. It supports multiple providers including GitHub Copilot, Claude Code, Google Gemini CLI, Codex, Cursor, and OpenCode. Give it a task, and a Project Lead agent breaks it down, assembles developers, architects, reviewers, and more, then coordinates their work in parallel while you stay in the loop.
npm install -g @flightdeck-ai/flightdeck
flightdeck
Instead of one AI agent doing everything sequentially, Flightdeck runs multiple agents at the same time — a developer writes code while a reviewer checks it, an architect designs the system, and a secretary tracks progress. The result: faster delivery, higher quality, and built-in checks and balances.
Lead Dashboard — your home screen for tracking projects, agents, and progress
Timeline — zoom, scroll, and replay agent activity with session scrubber
| Command Palette | Canvas View |
|---|---|
 |  |
| Mission Control | Analytics |
|---|---|
 |  |
| Timeline | Overview |
|---|---|
 |  |
| Batch Approval | New Project |
|---|---|
 |  |
npm install -g @flightdeck-ai/flightdeck
flightdeck
That's it. This installs Flightdeck globally and starts the server, opening the dashboard in your browser.
Example: "Refactor the auth module to use JWT tokens, add tests, and update the docs" → The lead creates a developer (implementation), a code reviewer (quality), and a tech writer (docs), sets up dependencies so the reviewer waits for the developer, and coordinates the whole flow.
CLI options: --port=4000 · --host=0.0.0.0 · --no-browser · -v / --version · -h / --help
npm install
npm run build
npm start
For development with hot reload:
npm run dev
http://localhost:3001http://localhost:5173 (dev) or http://localhost:3001 (production)DELEGATE/CREATE_AGENT@ in chat to autocomplete agent names; mentioned agents receive the message~ prefix and (est.) suffixreact-virtuoso, pinned user message banner, grouped sequential messages, per-project group chat historyprefers-reduced-motionPROGRESS command auto-reads DAG state, eliminating the need for separate queriesgit add only on files the agent has locked, then post-commit verification. Prevents git add -A from leaking other agents' work..github/skills/ with hot-reload via fs.watch, and AgentReconciliation on reconnectremember() / recall() for persistent agent knowledgeFlightdeck uses a three-tier architecture with clear separation of concerns:
┌──────────────────────────────────────────────────────────────────┐
│ Web Client (Vite/React) │
│ Dashboard · Project Views · Session Management · Real-time UI │
│ React 19 · TailwindCSS 4 · Zustand · WebSocket client │
└───────────────────────┬──────────────────────────────────────────┘
│ REST API + WebSocket (/ws)
┌───────────────────────▼──────────────────────────────────────────┐
│ Orchestration Server (Express 5) │
│ Session mgmt · Project mgmt · GovernancePipeline · Task DAG │
│ Knowledge · Decisions · File locks · Chat groups · Integrations │
│ SQLite/Drizzle · Zod validation · Pino logging │
└───────────────────────┬──────────────────────────────────────────┘
│ stdio + PID file (detached child process)
┌───────────────────────▼──────────────────────────────────────────┐
│ Agent Server (Daemon) │
│ Spawns & manages CLI agent processes via ACP protocol │
│ Copilot · Claude · Gemini · Codex · Cursor · OpenCode │
│ Per-agent lifecycle · Auto-restart · Heartbeat monitoring │
└──────────────────────────────────────────────────────────────────┘
Client ↔ Orchestration Server — The React frontend communicates with the Express server via REST endpoints (43 route modules covering agents, projects, sessions, tasks, decisions, knowledge, etc.) and a persistent WebSocket connection on /ws for real-time updates. WebSocket events are batched and throttled (agent text flushed every 100ms).
Orchestration Server ↔ Agent Server — The orchestration server forks the agent server as a detached child process, communicating via stdio and monitoring health with heartbeat pings. The agent server manages individual CLI agent processes (Copilot, Claude, Gemini, Codex, Cursor, OpenCode) through the ACP (Agent Client Protocol) — each agent runs as a separate subprocess with its own context window and role.
Command flow — Agents emit structured commands (wrapped in doubled Unicode brackets) in their output stream. The CommandDispatcher parses these and routes them to 10 domain-specific command modules (AgentCommands, TaskCommands, CommCommands, etc.). The GovernancePipeline intercepts commands through ordered hooks: security → permission → validation → rate-limit → policy → approval, with post-hooks for audit and metrics.
| Package | Description |
|---|---|
packages/shared | TypeScript types, Zod schemas, protocol definitions (shared between server and client) |
packages/server | Express 5 orchestration server, agent management, SQLite/Drizzle ORM, knowledge pipeline |
packages/web | React 19 + Vite frontend, Tailwind CSS 4, Zustand state management |
packages/docs | Documentation website |
| Component | Responsibility |
|---|---|
| AgentManager | Spawns agents, routes messages, manages delegations. 25+ typed events. Cascade termination with visited-set guard. |
| CommandDispatcher | Parses doubled Unicode-bracket commands (U+27E6/U+27E7) from agent output, routes to 10 command modules. |
| GovernancePipeline | Single interception point for all commands — pre-hooks (security, permission, validation, rate-limit, policy, approval) and post-hooks (audit, metrics). |
| TaskDAG + EagerScheduler | Directed acyclic graph for task scheduling with dependency resolution, parallel analysis, and eager pre-assignment of ready tasks. |
| ProjectRegistry | Persistent project management — CRUD, session tracking, briefing generation. |
| SessionResumeManager | Restores agent state after server restart with full context recovery. |
| KnowledgeInjector | Injects session knowledge, skills, and collective memory into agent context on spawn. |
| FileLockRegistry | Pessimistic file locking with TTL, glob support, expiry notifications. SQLite-backed. |
| MessageBus + ChatGroupRegistry | Inter-agent messaging, group chat lifecycle, role-based auto-membership. |
| DecisionLog | Decision tracking with accept/reject/reason workflow and approval queue. |
| ActivityLedger | Event-sourced activity logging (batched flushes every 250ms or 64 entries). |
| AlertEngine | Proactive detection: stuck agents, context pressure, duplicate edits, stale decisions. |
| IntegrationRouter | External platform routing — Telegram bot with batched notifications. |
| AdapterFactory | Selects agent backend (Copilot CLI, Claude SDK, Gemini, Codex, Cursor, OpenCode) based on configuration. |
See the Architecture Decisions page for the rationale behind key design choices.
Flightdeck uses layered configuration: hardcoded defaults ← YAML config ← environment variables ← runtime API.
Config file (flightdeck.config.yaml):
server:
maxConcurrentAgents: 50 # 1–200
heartbeat:
idleThresholdMs: 60000 # Idle agent detection threshold
crewUpdateIntervalMs: 180000 # Crew status push interval
models:
defaults: # Default model per role
lead: [claude-opus-4.6]
developer: [claude-opus-4.6]
architect: [claude-opus-4.6]
code-reviewer: [gemini-3-pro-preview]
# ... (14 roles total)
provider:
id: copilot # Active provider: copilot | claude | gemini | codex | cursor | opencode
budget:
limit: null # null = unlimited; set a dollar amount to cap spend
thresholds:
warning: 0.7
critical: 0.9
pause: 1.0
See flightdeck.config.example.yaml for the full reference.
Key environment variables:
| Variable | Default | Purpose |
|---|---|---|
PORT | 3001 | Server port |
HOST | 127.0.0.1 | Bind address |
DB_PATH | ./flightdeck.db | SQLite database location |
CLI_PROVIDER | copilot | Agent provider (copilot, claude, gemini, codex, cursor, opencode) |
ANTHROPIC_API_KEY | — | Required for Claude provider |
GEMINI_API_KEY | — | Required for Gemini provider |
OPENAI_API_KEY | — | Required for Codex provider |
TELEGRAM_BOT_TOKEN | — | Telegram integration (optional) |
AUTH | enabled | Set to none to disable authentication |
SERVER_SECRET | auto-generated | Fixed auth token (optional) |
MAX_AGENTS | 50 | Max concurrent agents (1–200) |
FLIGHTDECK_CONFIG | — | Path to YAML config file |
Each agent is assigned a role with a specialized system prompt. The lead creates agents and assigns them tasks.
| Role | Icon | Focus | Default Model |
|---|---|---|---|
| Project Lead | 👑 | Orchestration, delegation, crew coordination | Claude Opus 4.6 |
| Developer | 💻 | Code implementation, tests, bug fixes | Claude Opus 4.6 |
| Architect | 🏗️ | System design, technical debt, architecture decisions. Can delegate tasks. | Claude Opus 4.6 |
| Code Reviewer | 📖 | Readability, maintainability, code patterns | Gemini 3 Pro |
| Critical Reviewer | 🛡️ | Secure-by-design review, performance, edge cases | Gemini 3 Pro |
| Product Manager | 🎯 | User needs, product quality, UX review | GPT-5.3 Codex |
| Technical Writer | 📝 | Documentation, API design review, developer experience | GPT-5.2 |
| Designer | 🎨 | UI/UX, interaction design, accessibility | Claude Opus 4.6 |
| Generalist | 🔧 | Cross-disciplinary problem solving | Claude Opus 4.6 |
| Radical Thinker | 🚀 | Challenge assumptions, propose bold alternatives | Gemini 3 Pro |
| Secretary | 📋 | Plan tracking, status reports, session summaries | GPT-4.1 |
| QA Tester | 🧪 | Test strategy, quality assurance, coverage analysis | Claude Sonnet 4.6 |
| Agent | ⚙️ | Neutral general-purpose agent, no role-specific instructions | CLI default |
Custom roles can be created via the Settings UI with your own system prompts, colors, and icons.
AI-Aware Estimation: Traditional human timelines don't apply. A crew of AI agents can compress weeks of work into hours. Plan in sessions (30 min – 2 hours), not sprints. See the agent-collaboration-patterns skill for calibration data.
Agents communicate via structured commands wrapped in doubled Unicode brackets (⟦⟦ COMMAND {...} ⟧⟧, U+27E6/U+27E7) detected in their output stream. Commands are parsed by the CommandDispatcher and routed to the appropriate subsystem.
| Command | Description |
|---|---|
CREATE_AGENT {"role": "developer", "task": "..."} | Spawn a new agent with a specific role. Optionally assign a task and model. |
DELEGATE {"to": "agent-id", "task": "...", "context": "..."} | Assign a task to an existing agent. Leads and architects can delegate. |
TERMINATE_AGENT {"agentId": "agent-id", "reason": "..."} | Terminate an agent and free its slot. Logs session ID for potential resume. |
INTERRUPT {"to": "agent-id", "content": "..."} | Send a priority interrupt to a child agent, immediately stopping their current work. (Parent agents only) |
| Command | Description |
|---|---|
AGENT_MESSAGE {"to": "agent-id", "content": "..."} | Send a direct message to another agent by ID. |
DIRECT_MESSAGE {"to": "agent-id-prefix", "content": "..."} | Queue a message to another agent without interrupting their current work. Matches by ID prefix. |
BROADCAST {"content": "..."} | Send a message to all active agents. |
CREATE_GROUP {"name": "...", "members": ["id1"], "roles": ["developer"]} | Create a named chat group. Specify members by ID, by role, or both. Lead is auto-included. |
GROUP_MESSAGE {"group": "...", "content": "..."} | Send a message to all members of a group. Sender must be a member. |
ADD_TO_GROUP {"group": "...", "members": ["id"]} | Add agents to an existing group. New members receive recent message history. |
REMOVE_FROM_GROUP {"group": "...", "members": ["id"]} | Remove agents from a group. The lead cannot be removed. |
QUERY_GROUPS | List all groups the agent belongs to, with member counts and last message preview. |
QUERY_PEERS | Discover other active agents for direct messaging. |
REACT {"group": "...", "emoji": "👍"} | Add an emoji reaction to the latest (or specified) message in a group. |
| Command | Description |
|---|---|
DECLARE_TASKS {"tasks": [...]} | Declare a task DAG with dependencies. Tasks have taskId, role, dependsOn, and optional description. |
PROGRESS {"summary": "..."} | Report progress. Auto-reads DAG state when a DAG exists — no need to query separately. |
COMPLETE_TASK {"taskId": "task-id", "summary": "...", "output": "..."} | Mark a DAG task as done. Non-lead agents relay to parent's DAG with auth validation. Supports taskId, summary, status, output fields. (Any agent) |
TASK_STATUS | Query current task DAG status. |
PAUSE_TASK {"taskId": "..."} | Pause a pending/ready task in the DAG. (Lead-only) |
RETRY_TASK {"taskId": "..."} | Retry a failed task. (Lead-only) |
SKIP_TASK {"taskId": "..."} | Skip a task and unblock dependents. (Lead-only) |
ADD_TASK {"taskId": "...", "role": "...", "description": "...", "dependsOn": [...]} | Add a new task to an existing DAG. (Lead-only) |
CANCEL_TASK {"taskId": "..."} | Cancel a task. (Lead-only) |
RESET_DAG | Reset the entire DAG (clear all tasks). (Lead-only) |
DECISION {"title": "...", "rationale": "..."} | Log a decision. Users can accept/reject with a reason comment from the dashboard. |
QUERY_TASKS | Query current task DAG status (alias for TASK_STATUS). |
CANCEL_DELEGATION {"delegationId": "...", "reason": "..."} | Cancel an active delegation. |
ASSIGN_TASK {"taskId": "...", "agentId": "..."} | Assign a ready DAG task to an agent and move it to running state. (Lead-only) |
REASSIGN_TASK {"taskId": "...", "agentId": "..."} | Reassign a running task from one agent to another. (Lead-only) |
ADD_DEPENDENCY {"taskId": "...", "dependsOn": ["dep-id"]} | Add dependency edges to tasks in the DAG. Prevents circular dependencies. |
FORCE_READY {"taskId": "task-id"} | Force a pending/blocked task to ready state, overriding dependency checks. (Lead-only) |
| Command | Description |
|---|---|
LOCK_FILE {"filePath": "...", "reason": "..."} | Acquire a file lock. Prevents other agents from editing the same file. |
UNLOCK_FILE {"filePath": "..."} | Release a file lock. |
COMMIT {"message": "..."} | Scoped git commit — executes git add only on locked files, commits, then verifies files landed via git diff --name-only HEAD~1. Warns if expected files are missing. |
QUERY_CREW | Get the current roster of agents with IDs, roles, models, and status. |
ACTIVITY {"action": "...", "summary": "..."} | Log a structured activity entry to the activity ledger for auditing and tracking. |
HALT_HEARTBEAT | Pause automatic heartbeat nudges from the system. (Lead-only) |
REQUEST_LIMIT_CHANGE {"limit": 10, "reason": "..."} | Request to increase max concurrent agents. Requires user approval. (Lead-only) |
| Command | Description |
|---|---|
ACQUIRE_CAPABILITY {"capability": "code-review", "reason": "..."} | Temporarily gain capabilities beyond the agent's role (code-review, architecture, delegation, testing, devops). |
RELEASE_CAPABILITY {"capability": "code-review"} | Release a previously acquired capability. |
LIST_CAPABILITIES | List currently held capabilities. |
SET_TIMER {"label": "name", "delay": 300, "message": "...", "repeat": false} | Set a reminder that fires after a delay (in seconds). Optionally repeats. |
CANCEL_TIMER {"label": "name"} | Cancel an active timer. |
LIST_TIMERS | List all active timers. |
| Command | Description |
|---|---|
LIST_TEMPLATES | List all available workflow task templates with IDs, descriptions, and task sequences. |
APPLY_TEMPLATE {"template": "template-id"} | Instantiate a workflow template, creating tasks in the DAG. Supports overrides for customization. (Lead-only) |
DECOMPOSE_TASK {"task": "..."} | Break a task description into suggested sub-tasks with roles and dependencies. |
| View | Description |
|---|---|
| Lead Dashboard | Chat with the lead, decisions panel (accept/reject with reasons), crew/comms/groups/DAG/tokens tabs, three-tier message hierarchy, catch-up banner |
| Mission Control | Single-screen project overview: health summary, agent fleet, token economics, proactive alerts, activity feed, DAG minimap, comm heatmap, performance scorecards. Drag-and-drop panel reorder. |
| Agents | Unified list with hierarchy, model selector, plan progress, agent controls, project grouping |
| Tasks | Per-project task tabs with DAG status, progress badges, project grouping, duplicate detection |
| Timeline | Swim-lane visualization — filter by role/comm-type/status, brush time selector, keyboard navigation (←→ pan, +/- zoom), live auto-scroll mode, idle hatch patterns, hover tooltips |
| Group Chat | Tabbed group chat with human participation, project-level tab grouping, real-time messaging |
| Overview | Progress tracking, decision timeline grouped by project, global search |
| Settings | Concurrency limits (1–50 agents), model defaults, theme (Light/Dark/System), custom role editor, draggable dashboard panel layout |
busy_timeout, foreign_keys, synchronous=NORMAL)main and team-work-* branches — typecheck, unit tests, coverage upload| Document | Description |
|---|---|
| REST API Reference | Full REST API reference for all endpoints |
| Architecture Decisions | Key architecture decision records (ADRs) |
| Agent Communication | ACP agent communication protocol details |
| Coordination | File locking, delegation, and coordination primitives |
| Database Schema | SQLite schema and Drizzle ORM setup |
| UI Design | Frontend component architecture and design tokens |
Flightdeck is built on many excellent open-source projects. See docs/CREDITS.md for full attribution, including the research projects, agent SDKs, core libraries, and design references that made this possible.
FAQs
Flightdeck - AI crew management platform
The npm package @flightdeck-ai/flightdeck receives a total of 0 weekly downloads. As such, @flightdeck-ai/flightdeck popularity was classified as not popular.
We found that @flightdeck-ai/flightdeck demonstrated a healthy version release cadence and project activity because the last version was released less than a year ago. It has 1 open source maintainer collaborating on the project.
Threat Research
/Security News
Socket uncovered secret-stealing workflows that ran in LiteLLM’s GitHub repository 11 days before the supply chain attack on PyPI and Docker.
Security News
Cloudflare introduces tsrs, a Rust port of TypeScript’s compiler promising faster checks and lower memory use, with the TypeScript team’s help.
Security News
The PHP Foundation says Chainguard broke a security embargo ahead of critical RCE patches, as leaks disrupted the coordinated release.