Agency Orchestrator (ao) is a multi-agent workflow orchestrator: a one-sentence task description is decomposed, matched against a library of 276 Chinese role definitions plus the 191 English roles of Agency Agents, laid out as a DAG and run with independent steps in parallel, producing a complete deliverable such as a feasibility analysis, a technology comparison, a business plan or a code review. Workflows are YAML files with no code: each step names a role (loaded as the system prompt), a task template with variable substitution, its dependencies, an optional acceptance standard, and optional mechanical assertions on the output (file count, minimum bytes, regex matches, required strings) that trigger one targeted rework round before the step fails.
The engine is a Node package with a CLI, a web Studio, a desktop app, a programmatic API and an MCP server mode that exposes six tools (run, validate, list, plan and compose workflows, list roles) to Claude Code, Cursor and other MCP clients. It runs on 20+ model APIs (DeepSeek, Claude, OpenAI, Gemini, xAI, Moonshot, Zhipu, Qwen, Volcengine or any OpenAI-compatible endpoint), or without an API key by reusing a logged-in coding CLI (Claude Code, Codex, Copilot CLI, Antigravity CLI, OpenClaw, Hermes, CodeBuddy, Cline, OpenCode) or a local Ollama model. The project's own evaluation finds multi-agent output clearly ahead of a single prompt on DeepSeek-class models and behind it on small local models, so 70B or larger is recommended for Ollama.
Around the run it saves well-performing role line-ups as reusable teams, optimizes and versions prompts, attaches skills (methodology playbooks such as TDD or structured code review) to a step, resumes a run from a chosen step, reworks a step with conversational feedback, schedules runs that push to Feishu or DingTalk groups, renders a shareable report page and keeps a ledger of manual interventions to compute an AI autonomy rate. A materialize option writes the code blocks a workflow produces into real project files, and an install command loads the role library into coding tools (Claude Code, Cursor, Copilot, WorkBuddy) so a coding agent continues from the generated scaffold.
Features
- One-sentence compose: selects roles from the library, designs the DAG and writes a runnable workflow YAML from a task description, with provider and model overrides
- YAML workflows: steps with role, task template, variable passing, dependencies, concurrency, acceptance criteria and mechanical assertions, retried with exponential backoff and saved per step
- Role library: 276 Chinese roles plus 191 English roles bundled, multilingual packs (Korean, Russian, Portuguese, Indonesian, Arabic), private role directories and single custom roles
- Provider connectors: 20+ APIs including DeepSeek, Claude, OpenAI, Gemini, xAI, Moonshot, Zhipu and Qwen, any OpenAI-compatible endpoint, and 10 no-key paths through logged-in coding CLIs or Ollama
- Teams: save a role line-up from a workflow and apply it to new tasks, shared between the CLI and Studio
- Prompt Lab: optimize, test, version and compare prompts, with side-by-side scoring in Studio
- Skills: attach a methodology playbook to a step, from the bundled superpowers-zh set or a custom directory
- Resume and feedback: rerun from a chosen step, or send a step's role conversational revision notes
- Scheduling and push: run on a schedule and deliver results to Feishu or DingTalk groups
- Reports and ledger: a shareable report page per run, and a manual-intervention ledger that rolls runs, AI-completed steps and human time into an autonomy rate
- MCP server: six tools (run, validate, list, plan and compose workflows, list roles) for Claude Code, Cursor and other MCP clients
- Studio and desktop app: a web interface for composing, role selection, provider setup and 70 built-in templates, plus a desktop build
