GitHub - Xateh/maestro: A harness for precise, auditable agent workflows — compose the graph, drive your local CLI agents, gate and replay every handoff. Built on LangGraph.

Your agents, conducted.

Getting Started · Why · Usage · Dashboard · Architecture

Maestro is a harness for precise, auditable agent workflows, built on LangGraph. You declare the exact graph — which roles run, which model drives each, where the gates and loops sit — and Maestro conducts the coding CLIs already installed on your machine (claude, codex, gemini, copilot, antigravity, ollama) across it. Every handoff between roles is typed, scoped, recorded, and replayable: raw output stays on disk, only compact contracts pass downstream, and any run can be re-opened and audited later. No API keys, no vendor lock-in, no per-token billing you didn't sign up for.

The pipeline isn't the point — the precision is. plan → execute → review is just the stock graph. Compose your own: a one-shot task, a long-running multi-stage project, or an unattended background service that polls your tracker and runs the right workflow on a cadence (maestro serve). Right model per role, gates where you want them, loops that terminate, every run inspectable after the fact.

Demo

One prompt, three roles, hands off automatically — plan, execute, review:

flowchart LR
    prompt([prompt]) --> planner[planner]
    planner -- handoff --> executor[executor]
    executor -- handoff --> reviewer[reviewer]
    reviewer --> done([done])

    planner -.- p1["claude<br/>plans + architecture"]
    executor -.- e1["codex<br/>writes + edits code"]
    reviewer -.- r1["codex<br/>reviews + approves"]

That's the stock default workflow — one graph among many. Mix and match freely: swap any role to any provider (gemini for big-context research, ollama for fully local, …) in .maestro/config.json or live in the TUI, or reshape the graph itself (maestro workflow use) — add stages, gates, and bounded loops. You hold the score; only the instruments change.

Why Maestro

Most developers already know which model to reach for: Gemini for deep research and big-context reading, Claude for planning and architecture, Codex for writing and editing code. Maestro makes that instinct automatic.

Instead of wrapping LLM APIs in your own glue code, Maestro drives the CLI tools already on your machine as subprocesses, each with its own authenticated session. If you can run claude --version, you're ready.

Features

CLI agents, no API keys — Maestro runs the coding CLIs already on your machine. No new credentials to provision, no API wrapper to maintain.
Right model per role — assign any provider to any role: Gemini for research-heavy planning, Claude for architecture and review, Codex for workspace writes.
LangGraph-powered flow — roles are graph nodes, transitions are edges; no bespoke state-machine code to maintain.
Compact typed handoffs — only { role, provider, payload, log_path } objects pass between roles. Raw stdout (300–400 KB a step) stays on disk and is never re-sent as prompt context.
Dual-backend persistence — every task, step, and handoff lands in .maestro/maestro.db (SQLite, default) or a PostgreSQL database when DATABASE_URL=postgres://… is set. Logs stay on disk; the DB stores paths.
Visible agent panes (optional acceleration) — the zero-dependency terminal backend is the default; install herdr on your PATH and the engine auto-selects it to seat each step in a terminal pane, one tab per task. Tabs close on success, stay open while a task waits on you, and a resumed task picks up in the same tab. Tune with herdr.close_tab_on, or force the default with MAESTRO_BACKEND=terminal.
MCP server — eight tools expose Maestro state, task creation, and workflow validation to any MCP-compatible agent (Claude Code, Cursor, …). One .mcp.json entry, no other config.
Full-screen TUI — a resize-aware, keyboard-driven terminal UI: live task board with filter views, one-keystroke approve/deny/answer, full-screen provider and role editors, a settings editor, and a workflow graph screen that draws roles, handoff arrows, and event transitions as a grid. Pipes and scripts get the classic prompt-driven TUI automatically.
Preflight + receipts — maestro doctor checks Node, provider CLIs, herdr, and .maestro state before anything runs; every task run ends with a per-role summary (duration, output size, outcome).
Portable roles (MRC) — author a role once as a unit (.maestro/roles/*.md, a superset of the Claude Code subagent frontmatter), or point a workflow role at an existing .claude/agents/*.md subagent or skill via source; Maestro normalizes all three into one role. Per-role tools allowlists are enforced where the provider supports it (Claude --allowedTools/--disallowedTools) and advisory elsewhere. See docs/role-convention.md.
Workflow templates — maestro init --workflow extended|local|solo|triage|research scaffolds ready-made pipelines: an all-local Ollama setup, an executor-only fast loop, a single-role classifier (triage), and a gather→synthesize research flow.
Import/export bundles — package a workflow as a shareable bundle and import it elsewhere (with automatic backup of the existing workflow).
Security model — host commands off by default, network binaries hard-denied even when allowlisted, secrets stripped from subprocess env, MCP file access path-traversal-guarded.
Interactive web dashboard — maestro serve exposes a Linear-inspired browser UI at http://localhost:<port>/. Live-polls task state every 5 s, filter tabs (All / Running / Retrying / Completed), click any row for a detail panel, and trigger an orchestrator refresh — all without page reloads.
OpenTelemetry tracing — set OTEL_EXPORTER_OTLP_ENDPOINT to export traces, spans, and auto-instrumented http/pg calls via OTLP. Zero overhead when the variable is unset.
Linear integration — optional server mode polls Linear and dispatches issues automatically.

Getting Started

Requirements

Requirement	Notes
Linux or macOS	Windows is not supported (Maestro relies on unix domain sockets and bash-spawned agent runners). On Windows, use WSL2.
Node.js ≥ 22.13	Uses the built-in `node:sqlite` (`DatabaseSync`). Check with `node --version`.
herdr (optional)	Optional acceleration over the zero-dependency terminal-pane default. Install separately and the engine auto-selects it; `MAESTRO_BACKEND=terminal` forces the default.
Provider CLIs	At least one of `claude`, `codex`, `copilot`, `gemini`, `antigravity`, `ollama` — whichever you already have installed and authenticated. The default workflow uses `claude` (planner) and `codex` (executor + reviewer).

Installation

# Clone and install
git clone git@github.com:Xateh/maestro.git
cd maestro
npm install

# Verify
node bin/maestro.mjs status

Global install (optional):

npm link         # makes `maestro` available on PATH
maestro status

As a nested package (monorepo):

# From your project root
git clone git@github.com:Xateh/maestro.git
cd maestro && npm install && cd ..

# Add shim scripts to your root package.json:
# "maestro":     "node maestro/bin/maestro.mjs",
# "maestro:mcp": "node maestro/src/mcp/server.mjs"

Quick Start

# Initialize .maestro/ in your project (config, workflow, dirs) + optional setup wizard
cd /path/to/your/project
maestro init

# Or pick a workflow template:
#   extended — adds a read-only System Evaluator + an `evaluate` audit mode
#   local    — all roles on ollama, zero cloud
#   solo     — executor only, fastest loop
maestro init --workflow extended

# Preflight: node version, provider CLIs, herdr, state dir, workflow, db
maestro doctor

# Create and run a task (planner → executor → reviewer)
maestro task "Add a /healthcheck endpoint to the Express app"

# Planner only — read the plan before anyone touches code
maestro task --plan-only "Refactor the authentication module"

# Watch and steer from the terminal UI
maestro tui

# List tasks
maestro status

# Dump full JSON state for one task
maestro inspect 20260608-120000-add-healthcheck

A task that needs you — a question, an approval — parks in waiting_user and keeps its terminal tab open with the conversation intact. Answer with maestro message, maestro approve, or the TUI, and the pipeline resumes in the same tab, same context.

Usage

Run Modes

Mode	Flow	Command
`task`	planner → executor → reviewer	`maestro task "<prompt>"`
`plan-only`	planner only; stops at handoff	`maestro task --plan-only "<prompt>"`
`evaluate`	system evaluator only (extended template)	`maestro task --mode evaluate "<prompt>"`
server	polls Linear, auto-dispatches	`maestro serve [--config <path>] [--port <n>]`

CLI Commands

Command	Purpose
`maestro task [flags] "<prompt>"`	Create and run a task
`maestro run-task <id>`	Re-run or continue an existing task
`maestro status` / `maestro inspect <id>`	List tasks / dump full task JSON
`maestro tui`	Interactive terminal UI
`maestro init` / `maestro doctor`	Scaffold `.maestro/` / preflight checks
`maestro message · approve · deny`	Answer or decide a waiting task
`maestro approve-action · deny-action · run-action · edit-action`	Manage pending action requests
`maestro retry · cancel · mark-done · extend-timeout`	Task lifecycle controls
`maestro project <subcommand>`	Multi-task project commands (worktrees)
`maestro setup <subcommand>`	Configure providers, keys, and imports
`maestro workflow <subcommand>`	Workflow file commands
`maestro role list · show · lint`	Inspect portable role units (`.maestro/roles`, `.claude/agents`, skills)
`maestro import-agent <path>`	Convert a `.claude/agents` subagent into a native role unit
`maestro export` / `maestro import <bundle>`	Share workflows as bundles
`maestro serve [--config <path>] [--port <n>]`	Server mode (Linear polling)

Run maestro help <command> for flags and details, or see docs/cli.md for the full reference.

The full-screen TUI (maestro tui) gives you a live task board and a workflow graph screen (press 2) that draws roles, handoff arrows, and event transitions.

Providers

Default mapping: planner = claude, executor = codex, reviewer = codex.

Provider	CLI binary	Plays best at
`claude`	`claude`	Planning, architecture, nuanced review — strong reasoning and instruction-following
`codex`	`codex`	Execution and editing — tight workspace integration, file writes, shell commands
`gemini`	`gemini`	Research-heavy planning — large context window, web-grounded tasks
`copilot`	`copilot`	Optional; good for teams already in the GitHub ecosystem
`antigravity`	`antigravity`	Optional; bring-your-own CLI
`ollama`	`ollama`	Fully local, offline-capable models — privacy-sensitive or air-gapped work. See docs/local-llm.md

Assign a provider to each role: role → provider mapping lives in .maestro/workflow.json (the providers block in config.json only defines each provider's adapter, models, and aliases). Edit it live in maestro tui (full-screen role editor), or by hand:

// .maestro/workflow.json — defaults shown; swap any provider
{
  "roles": {
    "planner":  { "provider": "claude", "alias": "claude" },  // plans + architecture
    "executor": { "provider": "codex",  "alias": "codex"  },  // writes and edits code
    "reviewer": { "provider": "codex",  "alias": "codex"  }   // reviews and approves
  }
}

Want Gemini's large context for planning? Set planner.provider to gemini. Editing by hand replaces the whole file, so keep every role you want — the TUI is the safe path. See docs/configuration.md for the full role schema.

Terminal backend (the zero-dependency default):

MAESTRO_BACKEND=terminal maestro task "..."

The terminal backend runs agents via direct child_process.spawn (no visible panes) and is the default — a fresh install with nothing else on PATH uses it automatically (with a one-line notice). herdr is an optional acceleration: when its binary is found on PATH, the engine auto-selects it to seat each step in a visible terminal pane. Set MAESTRO_BACKEND=terminal to force the default even when herdr is installed. The terminal backend is covered by its own CI lane (npm run test:terminal).

Configuration

State and config live in .maestro/ in your working directory (or override with --state-dir):

.maestro/
  config.json       # providers, timeouts, planner policy, worktrees, tab lifecycle
  workflow.json     # roles, transitions, prompt templates
  maestro.db        # SQLite: tasks, steps, handoffs (LangGraph engine)
  tasks/            # legacy per-task JSON state
  runs/             # per-run logs: <role>.stdout.log, handoff.<role>.json
  projects/         # project state

See docs/configuration.md for the full schema — including herdr.close_tab_on ("success" | "terminal" | "never"), which decides when a task's terminal tab closes.

MCP Integration

Maestro exposes eight read/create/validate tools via MCP stdio transport.

Add to your .mcp.json:

{
  "mcpServers": {
    "maestro": {
      "command": "node",
      "args": ["/path/to/maestro/src/mcp/server.mjs"]
    }
  }
}

Tool	Purpose
`maestro_list_tasks`	List tasks, filter by status, newest-first
`maestro_show_task`	Task JSON + handoffs + stdout log tails
`maestro_list_runs`	Recent run directories
`maestro_show_run`	All files in one run
`maestro_create_task`	Spawn a new task by prompt
`maestro_get_state`	Runtime state snapshot (HTTP → file fallback)
`maestro_read_workflow`	Current `workflow.json` graph definition
`maestro_validate_workflow`	Validate a workflow definition before use

Full schema: src/mcp/SCHEMA.md · Extended docs: docs/mcp.md

Security Model

host_command off by default. Action requests that exec host commands are rejected at approval time unless .maestro/config.json has "host_command_allow": ["binary1", ...]. Network/privilege-escalation binaries (curl, wget, ssh, sudo, …) are hard-denied even if listed.
Env key denylist. LD_PRELOAD, PATH, GIT_SSH_COMMAND, NODE_OPTIONS, BASH_ENV, DYLD_*, GIT_PROXY* are stripped from all action-request env objects at parse time.
MCP path traversal guard. maestro_show_task and maestro_show_run reject IDs that do not match ^[0-9A-Za-z][0-9A-Za-z._-]{0,127}$ and verify the resolved path stays inside .maestro/.
Config redaction. maestro_get_state strips keys matching *_key/*_token/*_secret/api_key/apikey/password/passwd before returning config to MCP clients.
HTTP rate limiting + input validation. The dashboard/API server applies a per-IP token-bucket limit (429 + Retry-After when exceeded) and validates every route: identifiers are length-capped and charset-restricted, malformed encodings return 400, and oversized bodies return 413. Toggle with MAESTRO_HTTP_RATELIMIT.

Agents themselves run with your user's privileges — review what you approve. See SECURITY.md for the vulnerability reporting policy.

Web Dashboard

When running maestro serve, Maestro starts an HTTP server (default port from config.json → server.port). Visit http://localhost:<port>/ for the dashboard:

Live task board — auto-polls /api/v1/state every 5 s (active tasks) or 30 s (idle). Updates sidebar counts, task rows, and token totals in-place.
Filter tabs — All / Running / Retrying / Completed.
Detail panel — click any row to fetch /api/v1/<identifier> and inspect full issue data: state, attempt, timestamps, description, priority, assignee. Copy JSON to clipboard or open the raw endpoint in a new tab.
Trigger Refresh — sends POST /api/v1/refresh then polls immediately; button shows a spinner while in-flight.

The JSON API is also directly usable:

Endpoint	Method	Description
`/api/v1/state`	GET	Full orchestrator snapshot
`/api/v1/<identifier>`	GET	Runtime details for one issue
`/api/v1/refresh`	POST	Force an immediate tick + Linear reconcile

Observability

Maestro emits OpenTelemetry traces when OTEL_EXPORTER_OTLP_ENDPOINT is set. It is a no-op (zero overhead, zero imports) when the variable is absent.

# Export to a local Jaeger or Grafana Tempo collector
OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318 maestro serve

Auto-instrumented: http, pg (when using the PostgreSQL backend), dns. Service name defaults to the package name (maestro-orchestrator); override with OTEL_SERVICE_NAME.

Known Limitations

Maestro is a capable local orchestrator. The following areas are not yet implemented for production/enterprise deployments:

Authentication & RBAC: Single-user design; no identity provider integration or role-based access control.
Secrets Management: Relies on local CLIs for auth. No native HashiCorp Vault or AWS KMS integration.
Containerization: No official Dockerfiles or Kubernetes Helm charts.

Architecture

flowchart TD
    cli["bin/maestro.mjs<br/>CLI entry"] --> engine["LangGraph engine<br/>src/langgraph/"]
    engine --> nodes["role nodes<br/>planner · executor · reviewer"]
    nodes --> adapters["adapters<br/>src/adapters/"]
    adapters --> procs["provider subprocesses<br/>claude · codex · gemini · …"]

    engine -.-> db[("DB<br/>SQLite / Postgres")]
    nodes -.-> logs[("run logs<br/>.maestro/runs/")]
    procs -.-> herdr["herdr panes<br/>(default backend)"]

Full module tree

bin/maestro.mjs (CLI entry)
│
├─ src/langgraph/          LangGraph engine
│   ├─ engine.mjs          runLangGraphTask() — entry point
│   ├─ graph.mjs           buildGraph() from workflow.json
│   ├─ nodes.mjs           makeRoleNode() — node factory
│   ├─ prompt.mjs          compact prompt builder (typed handoffs)
│   └─ state.mjs           MaestroState channels
│
├─ src/adapters/           Provider command builders (pure functions)
│   ├─ registry.mjs        resolveAdapter() — built-in:<name> dispatch
│   └─ claude · codex · copilot · gemini · antigravity · ollama
│
├─ src/setup/              init scaffolding, doctor, import/export, templates
├─ src/db/
│   ├─ store.mjs           SqliteTaskStore (node:sqlite, default)
│   └─ pg-store.mjs        PostgresTaskStore (pg pool, activated by DATABASE_URL)
├─ src/telemetry.mjs       OpenTelemetry SDK init (no-op unless OTEL_EXPORTER_OTLP_ENDPOINT set)
├─ src/herdr-client.mjs    JSON-RPC wrapper around herdr binary
├─ src/herdr-agent-runner.mjs  HerdrAgentRunner (default backend, tab lifecycle)
├─ src/agent-runner.mjs    TerminalAgentRunner (fallback backend)
├─ src/orchestrator.mjs    MaestroOrchestrator (server mode + tick loop)
├─ src/router.mjs          buildStepPrompt, evaluatePlannerDecision
├─ src/state-machine.mjs   Pure transition(state, event) → nextState
├─ src/task-store.mjs      LocalTaskStore + DEFAULT_WORKFLOW
├─ src/setup/server-config.mjs  resolveServerConfig, renderPrompt (server mode)
├─ src/task-graph-runner.mjs    TaskGraphRunner (issue → graph task adapter)
├─ src/workflow-validate.mjs  Workflow schema validation
├─ src/workspace.mjs       WorkspaceManager (git worktrees)
├─ src/markers.mjs         Pure parsers: HANDOFF/QUESTION/REVIEW/ACTION_REQUEST
├─ src/logger.mjs          StructuredLogger (logfmt, crash-safe)
├─ src/http-server.mjs     HTTP server: interactive dashboard (/) + JSON API (/api/v1/*)
├─ src/linear-tracker.mjs  Linear GraphQL issue fetcher
├─ src/tui.mjs + src/tui/  Interactive terminal UI (full-screen + classic)
└─ src/mcp/server.mjs      MCP stdio server (8 tools)

Full architecture documentation: docs/architecture.md

Environment Variables

Environment variables (full table)

Variable	Default	Description
`MAESTRO_BACKEND`	auto (herdr when installed)	`"terminal"` to bypass herdr panes; without herdr on PATH, terminal is used automatically
`MAESTRO_ROOT`	cwd walk	Override runtime root (where `.maestro/` lives)
`MAESTRO_TUI_CLASSIC`	unset	`1` forces the classic prompt-driven TUI
`HERDR_BIN`	`"herdr"`	Path to the herdr binary
`HERDR_SOCKET_PATH`	`~/.config/herdr/herdr.sock`	herdr daemon socket
`DATABASE_URL`	unset	PostgreSQL connection string (`postgres://user:pass@host/db`). When set, Maestro uses PostgreSQL instead of SQLite for all task/handoff state.
`OTEL_EXPORTER_OTLP_ENDPOINT`	unset	OTLP collector base URL (e.g. `http://localhost:4318`). Enables OpenTelemetry tracing. No-op when unset.
`OTEL_SERVICE_NAME`	`maestro-orchestrator`	Override the OTel service name reported in traces.
`MAESTRO_SECRET_PASSPHRASE`	unset	Unlocks the encrypted secret store (`secrets.local.enc.json`) without an interactive prompt. Real environment variables still take precedence.
`MAESTRO_OLLAMA_BIN`	`"ollama"`	Ollama binary or alias used by the `agent:ocr`/`agent:eval` helper scripts.
`MAESTRO_OLLAMA_MODEL`	`"llama3.2"`	Text model for the `agent:eval` helper script.
`MAESTRO_OLLAMA_VISION_MODEL`	`"llama3.2-vision"`	Vision model for the `agent:ocr` helper script (multimodal).
`HEADROOM_PROXY_URL`	`http://localhost:8787`	Headroom compression proxy endpoint used for prior-output context compaction.
`NO_COLOR`	unset	Set to any value to disable ANSI color in CLI/TUI output (honors the NO_COLOR standard).
`MAESTRO_HTTP_RATELIMIT`	on	Set to `off` to disable the dashboard/API per-IP rate limiter (reads ~120/min, writes ~12/min; `429` + `Retry-After` otherwise).

Development

npm test                # full suite — node --test, hermetic
npm run lint            # Biome (lint-only)
npm run test:coverage   # c8 coverage report
npm run test:enterprise # package + maestro + mcp suites in sequence

Tests use the Node.js built-in test runner. No API keys, no agent CLIs, no herdr binary needed — stub runners and temp dirs all the way down. CI runs lint, the test matrix (Node 22 and 24), coverage, and npm audit on every push and pull request.

Releases are manual and documented in RELEASING.md.

Documentation

docs/architecture.md — module index, run flow, SQLite schema, extension recipes
docs/cli.md — every command, flags, env vars
docs/configuration.md — .maestro/ layout, config schema, provider setup, tab lifecycle
docs/import-export.md — workflow bundle format, export/import flow
docs/role-convention.md — portable role units (MRC), tool allowlists, provider capability matrix
docs/local-llm.md — fully local setup with Ollama
docs/mcp.md — MCP tools reference and registration
src/mcp/SCHEMA.md — raw MCP tool schema (kept in sync with server.mjs)
CHANGELOG.md — release history (Keep a Changelog format)

Contributing

Contributions are welcome — see CONTRIBUTING.md for the workflow, code style (Biome), and test expectations. Security issues should follow the process in SECURITY.md.

Credits

See CREDITS.md for full acknowledgements: herdr · OpenAI Swarm (inspiration) · LangGraph · MCP SDK · OpenAI Codex CLI · LiquidJS · yaml

Name		Name	Last commit message	Last commit date
Latest commit History 141 Commits
.github		.github
.maestro/roles		.maestro/roles
bin		bin
docs		docs
examples		examples
schema		schema
scripts		scripts
src		src
test		test
.c8rc.json		.c8rc.json
.editorconfig		.editorconfig
.gitattributes		.gitattributes
.gitignore		.gitignore
.nvmrc		.nvmrc
CHANGELOG.md		CHANGELOG.md
CODE_OF_CONDUCT.md		CODE_OF_CONDUCT.md
CONTRIBUTING.md		CONTRIBUTING.md
CREDITS.md		CREDITS.md
LICENSE		LICENSE
README.md		README.md
RELEASING.md		RELEASING.md
SECURITY.md		SECURITY.md
biome.json		biome.json
package-lock.json		package-lock.json
package.json		package.json

Uh oh!

Folders and files

Latest commit

History

Repository files navigation

Demo

Table of Contents

Why Maestro

Features

Getting Started

Requirements

Installation

Quick Start

Usage

Run Modes

CLI Commands

Providers

Configuration

MCP Integration

Security Model

Web Dashboard

Observability

Known Limitations

Architecture

Environment Variables

Development

Documentation

Contributing

Credits

License

About

Resources

License

Code of conduct

Contributing

Security policy

Uh oh!

Stars

Watchers

Forks

Releases 8

Sponsor this project

Uh oh!

Packages 0

Uh oh!

Uh oh!

Contributors

Uh oh!

Languages

Packages