M
MCP Directory CURATED DIRECTORY
911 MCP servers indexed
中文
← Back to directory
A

Agency Orchestrator

Community
One sentence to assemble your AI expert team, deliver complete deliverables in minutes.
How to install ↓ Compare GitHub source repository ↗
Category
Dev Tools #74 of 438
Stars
★ 2.4k Very popular
Transport
stdio (local process)
Runtime
Node.js
Credentials
Optional API key
License
Apache-2.0
Last commit
Tools
6
63FMRS · C

Agency Orchestrator is feature-rich and suitable for quickly building multi-agent collaboration workflows. Installation and configuration are simple, and the MCP interface facilitates integration. However, attention is needed regarding LLM capabilities and data security risks.

Strongest · Maintenance 16/20 Weakest · Reliability 10/20

Reliability
10/20
Security and permissions
12/20
Maintenance
16/20
Documentation
15/20
Setup experience
10/20
Why each score
Reliability 10/20
README shows a clear CLI/MCP architecture with `ao serve` for MCP and six tools. There is a contradiction: README claims 'npm + 2 dependencies' but actual project may have more. No tests or CI workflows were evident in the static review. Error handling and edge cases are thinly documented. Reliability is capped at 10 due to lack of verifiable execution evidence.
Security and permissions 12/20
API keys are handled via environment variables, no hardcoded keys. Several keyless providers (e.g., claude-code, gemini-cli) reduce key exposure, but invoking external LLMs via CLI sends code/text to third parties; data flow is disclosed but cross-network boundaries are not fully clarified. No malicious behavior found. Approval nodes (type: approval) require human confirmation, but MCP tools like run_workflow may execute workflows by default, including steps with external side effects. No explicit confirmation is evident. Security score 12, no red line issues.
Maintenance 16/20
Apache-2.0 license is clear, CI badge exists (though CI file not seen in static review), npm version suggests active releases, repository is active (README mentions v0.5, roadmap v0.6). No security response channel identified, but common for Apache projects. Dependency updates and issue response not directly confirmable. Maintenance score 16.
Documentation 15/20
README is extremely thorough: installation, configuration, 11 LLM providers, CLI commands, YAML schema, MCP config examples, built-in workflows, output structure. But lacks dedicated detailed documentation for MCP tools (parameters, error codes, limits). Cost is mentioned (e.g., DeepSeek top-up) but no concrete pricing. Limitations are disclosed (quality of local small models). Overall documentation good, but MCP tool docs thin, score 15.
Setup experience 10/20
Installation is straightforward: `npm install -g agency-orchestrator`, then `ao serve` for MCP. Configuration examples for Claude Code and Cursor are provided. However, dependency list is vague (claims 2 but actual may be more), and using a provider requires either a local CLI (like Claude Code) or API key setup, adding complexity. Docker deployment instructions exist. No execution verification possible, so actual setup may have surprises; score 10 (less than 15 cap due to no verifiable evidence).

Static review · not runListed 2026-08-07

Read the FMRS scoring method →

Fit and risk

What it can accessWrites / deletes local filesRuns commands or codeUses the network

Best for

  • Developers and creators who want to use AI multi-agent collaboration for complex tasks (like business planning, technical solutions, content creation).
  • Teams that prefer no-code YAML workflows over writing Python code.
  • Users who want to leverage existing AI subscriptions (like Claude, Gemini, Copilot) to reduce costs.

Not for

  • Deep customization scenarios requiring fine-grained control over each agent's internal logic.
  • Enterprises with strict data privacy requirements, as some API calls require network access.
  • Scenarios relying on small local models (e.g., 8B level) for high-quality outputs, as multi-role handoffs may amplify drift.

Required permissions

  • When executing workflows, it calls configured LLM APIs or local CLIs, requiring network access and API permissions.
  • It can read/write local file system (e.g., read input files, save outputs to `ao-output/` directory).
  • The MCP server itself only provides stdio communication and does not actively access external systems.

Risks and side effects

  • LLM outputs may be inaccurate or hallucinated; acceptance criteria can mitigate but not fully avoid.
  • When using third-party APIs, ensure API keys are securely stored to prevent leakage.
  • Workflows with human approval nodes require user intervention, otherwise the flow pauses.

Setup

Before you start

Runtime：Node.js

DEEPSEEK_API_KEY optionalsecret API key from the DeepSeek open platform; one key-based option, not required since key-free CLI providers exist.
ANTHROPIC_API_KEY optionalsecret Claude API key from the Anthropic console; needed only when provider is claude.
OPENAI_API_KEY optionalsecret OpenAI or any OpenAI-compatible API key; needed when provider is openai.
ARK_API_KEY optionalsecret Volcengine (Doubao/Kimi/GLM) API key from the Volcengine Ark console.
Other optional settings (10)
OPENAI_BASE_URL optional Custom OpenAI-compatible endpoint (e.g. Zhipu, SiliconFlow, Moonshot), set in .env.
AO_PROVIDER optional Default LLM provider set in .env, e.g. openai.
AO_MODEL optional Default model name set in .env.
AO_DATA_DIR optional Custom output/data directory; desktop defaults to userData.
AO_AGENTS_DIR optional Custom role library directory; run/compose/roles/web all switch to it when set.
AO_HOME optional Unified workspace root (e.g. ~/.ao) keeping outputs and generated workflows in one place.
AO_OUTPUT_DIR optional Separately specify the run-output directory.
AO_WORKFLOWS_DIR optional Separately specify the workflows directory.
AO_PROMPTS_DIR optional Prompt asset storage directory, default ~/.ao/prompts/.
AO_SKILLS_DIR optional Custom skills directory; defaults to the bundled superpowers-zh.
  1. Install the CLI globally: npm install -g agency-orchestrator.
  2. Start the MCP server: ao serve.
  3. Configure in Claude Code's settings.json:
{
  "mcpServers": {
    "agency-orchestrator": {
      "command": "npx",
      "args": ["agency-orchestrator", "serve"]
    }
  }
}
  1. Or configure the same in Cursor's .cursor/mcp.json.
  2. Depending on the chosen LLM provider, set up API keys or use already-installed CLIs (e.g., Claude Code, Gemini CLI).
claude_desktop_config.json 
{
  "mcpServers": {
    "agency-orchestrator": {
      "command": "npx",
      "args": [
        "agency-orchestrator",
        "serve"
      ]
    }
  }
}

Shown for Claude Desktop. Other clients may use a different file or key (VS Code uses "servers") — the configurator below converts it.

.vscode/mcp.json
{
  "servers": {
    "agency-orchestrator": {
      "command": "npx",
      "args": [
        "agency-orchestrator",
        "serve"
      ]
    }
  }
}

Goes in your project's .vscode/mcp.json (VS Code uses a "servers" key).

Terminal
claude mcp add agency-orchestrator -- npx agency-orchestrator serve

Run it in a terminal; replace any <…> placeholders with your own values first.

Generate and validate client config

Check that it works

After adding npx agency-orchestrator serve to your Claude Code or Cursor MCP config, the tool list should show run_workflow, validate_workflow, list_workflows, plan_workflow, compose_workflow, and list_roles; calling list_roles and getting roles back proves the connection.

Troubleshooting

  1. Ensure Node.js and npm are correctly installed, and agency-orchestrator is installed globally.
  2. Check the MCP configuration for correct command and arguments, especially the npx path.
  3. If using CLI as the LLM provider, ensure the corresponding CLI is installed and logged in.
  4. Verify the workflow YAML syntax, which can be checked with `ao validate`.
  5. Check terminal output or logs for specific error messages.

Things to try

Once connected, you can ask your AI assistant things like:

  • List all available AI roles and what each specializes in
  • Use compose_workflow to build a PR code review workflow covering security and performance from one sentence
  • Validate the workflows/dev/pr-review.yaml workflow configuration without running it
  • Run the one-person company all-hands workflow with the idea: use AI to help small merchants make short videos

Tools 6

run_workflow writes
Run a specified workflow with input variables, returning execution results.
validate_workflow read-only
Validate a workflow YAML configuration without executing it.
list_workflows read-only
List available workflow templates.
plan_workflow read-only
View the execution plan (DAG) of a workflow.
compose_workflow writes
Automatically orchestrate and generate a workflow YAML based on a natural language description.
list_roles read-only
List all available roles.

Use cases

Invoke workflows directly from your IDE via MCP for code review, requirements analysis, etc.
Describe a task in natural language and let AI automatically orchestrate a multi-role collaboration workflow.
Integrate automated workflows in CI/CD pipelines, such as PR review, security audit, etc.

Supported clients

Claude Desktop
Claude Code
Cursor

Listed from the project's documentation, not tested by this site.

Overview

Agency Orchestrator is an AI agent orchestration engine that defines multi-role collaboration workflows via YAML or natural language, automatically forming professional teams (267 Chinese roles + 184 English roles + 5 other languages), executing in parallel based on DAG, with automatic acceptance verification and human approval. Available via Web Studio, Desktop, Docker, or CLI, supporting 11 LLM providers (7 of which require no API key). Provides MCP Server mode for direct invocation by AI coding tools like Claude Code and Cursor.

Similar servers

Context7 80 · B

Upstash's official server providing up-to-date third-party library docs for AI coding assistants

★ 63.3k · Tools 2 Compare with this →
Compare with other MCP servers

Source revision 28da7ef7c177 Data synced 2026-10-19 Read the FMRS scoring method