Agency Orchestrator: One Sentence to Orchestrate an AI Expert Team
Turn one sentence into a 'one-person company' AI expert team delivering complete deliverables in minutes via zero-code YAML.
Evidence shows: CLI supports --resume and --from for rollback; approval step type for human confirmation; docs state API keys stay local, Web Studio does not upload; automatic backup and restore for Claude config. Deductions: least privilege not explicit, data flow transparency lacks detailed logs, sensitive data handling (e.g., key storage) insufficient, dependency security no audit or vulnerability scan evidence, external effects (e.g., network requests) not clearly stated.
Evidence shows: extensive tests (test/ directory) covering acceptance, azure-compat, etc.; dependencies listed in package.json but availability not guaranteed; failure messages have retry and timeout mechanisms. Deductions: dependency availability not verified, specific failure message content not shown.
Evidence shows: supports multiple LLM providers and languages, multiple usage modes (Web Studio, CLI, Docker), broad scenarios. Deductions: capability boundaries (e.g., role count limits) not explicit, trigger precision (e.g., compose accuracy) not verified, environment fit (e.g., different OS) not fully documented.
Evidence shows: README, LICENSE, package.json present, clear information architecture; detailed install notes; stable naming (ao command); examples and FAQ (tutorial link); known limitations (e.g., weak models) mentioned; Apache-2.0 license; version number exists but no changelog; maintenance responsibility not explicit. Deductions: no changelog, maintenance responsibility not explicit.
Evidence shows: output is complete plan, directly usable; high marginal value (multi-role collaboration); good cost-benefit (no API key options). Deductions: output usability not actually verified, cost-benefit not quantified.
Evidence shows: some claims in README supported by tests (e.g., acceptance test), but no independent verification; EVAL_FINDINGS.md mentioned but details not provided; fact vs inference separation unclear. Deductions: cross-source corroboration insufficient, fact vs inference separation unclear.
- Publisher identity unverified, proceed with caution.
- Dependency security lacks audit or vulnerability scan evidence; consider checking dependency versions.
- Data flow transparency and sensitive data handling details insufficient; review source code for key handling.
- No changelog, version history opaque.
What does this agent do, and when should you use it?
Agency Orchestrator is an open-source multi-agent orchestration engine that lets users describe a task in one sentence or define it in a YAML workflow, automatically assembling a team from 267 Chinese and 184 English professional AI roles (plus five additional language packs: Korean, Russian, Portuguese-BR, Indonesian, Arabic). It runs the team as a DAG, executes steps in parallel, and produces acceptance-ready outputs. The tool supports 11 LLM providers, seven of which require no API key (Claude Code, Gemini CLI, Copilot CLI, Codex CLI, OpenClaw, Hermes, Ollama), while also supporting API-key providers like DeepSeek, OpenAI, and Volcano Engine. Delivery options include a CLI, a graphical Web Studio, a cross-platform Electron desktop app, and Docker. It ships an MCP server that integrates with 14 AI coding tools including Claude Code and Cursor. Workflows support dependencies, conditions, loops, human-approval nodes, resumable runs, and conversational rework; every run saves a full output report with metadata.
The user issues a task via ao compose "one sentence" --run, or through the Web Studio, and the engine automatically decomposes the task, selects roles from the library, designs a DAG, generates the workflow YAML, and executes it. Each step is run by a designated role (e.g., product/product-manager), transfers outputs via {{variables}}, supports parallel execution (concurrency), dependencies (depends_on), conditions, loops, and approval/human-input node types. Results are saved under ao-output/ with summary.md, step files, and metadata.json (duration, token usage). Users can also run existing workflows with ao run workflow.yaml, and use ao serve to start an MCP server offering tools like run_workflow, validate_workflow, list_workflows, plan_workflow, compose_workflow, and list_roles, which can be called from Claude Code or Cursor. The Web Studio provides a visual canvas for drag-and-drop workflow editing, real-time progress highlighting, multi-language role library switching, and prompt optimization.
- An entrepreneur uses
ao composeto generate a feasibility analysis report, automatically grouping market, finance, and product roles to work in parallel. - A developer runs a PR review workflow on CI with
ao run, invoking code review, security, and performance roles. - A content creator uses the built-in
story-creation.yamltemplate for collaborative novel writing, using--feedbackfor conversational rework. - A tech lead uses
ao planandao explainto preview the execution plan and understand the multi-agent collaboration before running. - Someone with an existing Claude Pro subscription uses
--provider claude-codewith zero configuration, no API key needed. - A team saves a proven role lineup via
ao team saveand reuses it on new tasks withao run --team <name> "new task".
What are this agent's strengths and limitations?
- Supports 11 LLM providers, with 7 requiring no API key (Claude Code, Gemini CLI, Copilot, Codex, Ollama, etc.), dramatically lowering the entry barrier.
- Provides 267 Chinese and 184 English roles covering product, engineering, design, and marketing, ready to use.
- Zero-code YAML workflow definition with automatic DAG construction, variable passing, conditionals, and loops, reducing orchestration complexity.
- Offers a graphical Web Studio, desktop app, and Docker deployment to suit different user preferences.
- Integrates with 14 AI coding tools via MCP server, enabling use within Claude Code, Cursor, and others.
- Includes 32 built-in workflow templates for development, marketing, strategy, and HR, ready to run.
- Lightweight local models (like 8B) are not strong enough and may amplify drift in multi-agent collaboration, requiring capable models like DeepSeek/Claude/Gemini for best outcomes.
- CLI requires Node.js and global npm install; the desktop app is heavier though it avoids npm.
- Some providers (e.g., Gemini CLI, Codex CLI) require installing and logging into additional CLI tools.
- API-key providers (e.g., DeepSeek) incur usage costs, though they are relatively low.
- Customizing YAML workflows requires learning the schema, which has a learning curve.
How do you install or deploy this agent?
Install the CLI globally: npm install -g agency-orchestrator. For the desktop app, download the latest Electron release from GitHub (macOS, Windows, Linux); it bundles the engine and Node, no npm needed. Docker deployment: docker run -d -p 8088:8088 -v ao-data:/data ghcr.io/jnmetacode/agency-orchestrator:latest or use the provided docker-compose.yml. First run often needs no configuration: if Claude Code or Gemini CLI is already logged in, the tool auto-detects and uses it. To use DeepSeek, set the environment variable DEEPSEEK_API_KEY or use ao init. For other API-key providers, configure via ao init --provider openai --model ... --base-url ... --api-key ... or edit .env.
How do you use this agent?
Basic usage: ao compose "analyze the feasibility of an AI bookkeeping tool" --run --provider deepseek (or --provider claude-code if you have a Claude subscription). The engine will generate and run the workflow, saving outputs to ao-output/. To use the Web Studio, run ao web and open the browser to visually select experts, run workflows, and inspect artifacts. You can also run a built-in template directly: ao run workflows/one-person-company-all-hands.yaml --input idea="...". To use with Cursor or Claude Code, set up the MCP server (ao serve) and use the provided tools. Other CLI commands include ao roles, ao init --lang en, ao team save/list, ao prompt optimize, and more.
How does this agent compare with similar options?
Compared to ChatGPT/Claude, this tool provides multi-role parallel collaboration instead of single-turn conversation. Compared to CrewAI/LangGraph (Python-based), this tool requires no coding, is npm-installable, and uses YAML configuration, bundling over 400 roles.
FAQ
Which model providers incur extra costs?
How does the tool ensure output quality?
acceptance criterion that gets injected into the prompt and used for review. You can also add human-approval nodes to gate critical decisions. Official evaluation recommends using models at the DeepSeek level or above.What happens if a step fails?
llm.retry, resume from the last run with --resume last, or re-run from a specific step with --from <step-id>.Can it run offline?
How can I add custom roles?
AO_AGENTS_DIR environment variable to point to your own role directory, or place role files in ~/.ao/roles/ and see them in the Studio under 'My Roles'.