Conductor

A CLI tool for defining and running multi-agent workflows with the GitHub Copilot SDK and Anthropic Claude.

Why Conductor?

Conductor makes multi-agent workflows — code review pipelines, research-then-synthesize flows, plan-then-implement loops — repeatable, deterministic, and version-controlled. You define your agents, their prompts, and the routing between them in a single YAML file:

• Repeatable — Same inputs follow the same path through the same agents.
• Deterministic — Routing uses Jinja2 templates and expression evaluation. First matching condition wins. No LLM in the orchestration loop, no tokens spent deciding what runs next.
• Source-controlled — Plain YAML files. Diff workflows in pull requests, version them with your code, run them the same way locally and in CI.

Features

• YAML-based workflows - Define multi-agent workflows in readable YAML
• Multiple providers - GitHub Copilot, Anthropic Claude, Claude Agent SDK (experimental), NousResearch Hermes (experimental), or Azure Container Apps sandboxed execution (experimental) with seamless switching
• Parallel execution - Run agents concurrently (static groups or dynamic for-each)
• Sub-workflow composition - Reusable sub-workflows with templated input_mapping, usable inside for_each groups for dynamic fan-out
• Script steps - Run shell commands and route on exit code or parsed JSON stdout
• Set steps - Bind one or more Jinja2-evaluated values into the context (no LLM, no subprocess) for derived flags, computed defaults, and constants reused by many later prompts
• Terminate steps - Explicit terminal step with status (success/failed) and structured reason — distinguishable from the default $end path in CLI exit codes, dashboard state, and event logs
• Dialog mode - Agents can pause for multi-turn conversation when uncertain
• Reasoning effort - Unified reasoning.effort (low/medium/high/xhigh/max) per agent or workflow-wide, translated to each provider's native API
• Workspace instructions - Auto-discover and inject AGENTS.md / CLAUDE.md / .github/copilot-instructions.md into every agent's prompt
• Conditional routing - Route between agents based on output conditions
• Human-in-the-loop - Pause for human decisions with Markdown-rendered prompts and clickable file links
• Safety limits - Max iterations and timeout enforcement
• Web dashboard - Real-time workflow visualization with interactive DAG graph, breadcrumb navigation into sub-workflows, live streaming, and in-browser human gates
• Validation - Catches stale template references, missing inputs, and undeclared dependencies before runtime

Installation

Quick Install (Recommended)

macOS / Linux:

curl -sSfL https://aka.ms/conductor/install.sh | sh

Windows (PowerShell):

irm https://aka.ms/conductor/install.ps1 | iex

The installer checks for uv (installs it if missing), fetches the latest release with pinned dependencies, and verifies integrity via SHA-256 checksum.

Updating

conductor update checks for a newer release and tells you the one-line command to upgrade. Upgrades happen via the install script — the same script you used to install — because in-process self-upgrade is unreliable on Windows (the running Python interpreter sits inside the venv that needs replacing).

conductor update

To upgrade, run the install script in a new shell (not from inside a running conductor process):

macOS / Linux:

curl -sSfL https://aka.ms/conductor/install.sh | sh

Windows (PowerShell):

irm https://aka.ms/conductor/install.ps1 | iex

Or skip the copy-paste with --apply:

conductor update --apply

--apply launches the install script automatically — on Windows it opens in a new console window so you can watch progress; on macOS/Linux it replaces the current process. Either way, the running conductor exits before the installer touches the venv, so file locks release cleanly.

The install script handles file-lock safety (process detection, stale-file cleanup, and on Windows a rename-fallback when the venv directory can't be removed), retries with backoff, and verifies the installed version after install. If your shell ever gets into a bad state from a failed update, re-running the install script is always the right next step.

Conductor periodically checks GitHub for newer releases (cached for 24 hours under ~/.conductor/update-check.json) and prints a one-line hint when one is available. To silence the hint permanently — for example when you manage upgrades through a package manager or company-mirrored install — set CONDUCTOR_NO_UPDATE_CHECK=1 in your shell environment. The check is also skipped automatically for non-TTY invocations, --silent mode, the update subcommand, and --help / --version.

Manual Install

# Install from GitHub
uv tool install git+https://github.com/microsoft/conductor.git

# Run the CLI
conductor run workflow.yaml

# Or run directly without installing
uvx --from git+https://github.com/microsoft/conductor.git conductor run workflow.yaml

# Install a specific branch, tag, or commit
uv tool install git+https://github.com/microsoft/conductor.git@branch-name
uv tool install git+https://github.com/microsoft/conductor.git@v1.0.0
uv tool install git+https://github.com/microsoft/conductor.git@abc1234

Using pipx

pipx install git+https://github.com/microsoft/conductor.git
conductor run workflow.yaml

# Install a specific branch or tag
pipx install git+https://github.com/microsoft/conductor.git@branch-name

Using pip

pip install git+https://github.com/microsoft/conductor.git
conductor run workflow.yaml

# Install a specific tag or commit
pip install git+https://github.com/microsoft/conductor.git@v1.0.0

Use the Conductor skill in Claude Code or Copilot CLI

This repo doubles as a single-plugin marketplace that ships the conductor skill from plugins/conductor/skills/conductor/. The skill teaches the assistant the workflow YAML schema, CLI commands, and execution model.

Claude Code:

/plugin marketplace add microsoft/conductor
/plugin install conductor@conductor

GitHub Copilot CLI (gh skill requires GitHub CLI 2.91+, public preview):

gh skill install microsoft/conductor conductor

The plugin ships only markdown — no executables, hooks, or MCP servers — so trust verification is straightforward.

Quick Start

1. Create a workflow file

# my-workflow.yaml
workflow:
  name: simple-qa
  description: A simple question-answering workflow
  entry_point: answerer

agents:
  - name: answerer
    model: gpt-5.5
    prompt: |
      Answer the following question:
      {{ workflow.input.question }}
    output:
      answer:
        type: string
    routes:
      - to: $end

output:
  answer: "{{ answerer.output.answer }}"

2. Run the workflow

conductor run my-workflow.yaml --input question="What is Python?"

3. View the output

{
  "answer": "Python is a high-level, interpreted programming language..."
}

Web Dashboard

Conductor includes a built-in real-time web dashboard that lets you visualize and interact with your workflows as they run. Launch it with --web:

conductor run workflow.yaml --web --input question="What is Python?"

Key features:

• Interactive DAG graph — Zoomable, draggable workflow graph with animated edges showing execution flow and conditional routing
• Live agent streaming — Watch agent reasoning, tool calls, and outputs stream in real-time as each step executes
• Three-pane layout — Resizable panels for the graph, agent detail, and a tabbed output pane (Log, Activity, Output)
• In-browser human gates — Respond to human-in-the-loop decision points directly in the dashboard, no terminal needed
• Per-node detail — Click any node to see its prompt, metadata (model, tokens, cost), activity stream, and output
• Background mode — Run with --web-bg to start the dashboard in the background, print the URL, and exit. Use conductor stop to shut it down later.

# Run in background — prints dashboard URL and exits
conductor run workflow.yaml --web-bg --input topic="AI in healthcare"

# Stop a background workflow
conductor stop

Providers

Conductor supports multiple AI providers. Choose based on your needs:

Feature	Copilot	Claude	Claude Agent SDK	Hermes	ACA
Tier	Stable	Stable	Experimental	Experimental	Experimental
Pricing	Subscription	Pay-per-token	Subscription	Pay-per-token (via hermes)	Subscription + ACA compute
Context Window	Per-model	Per-model	Per-model	Per-model	Per-model (inner Copilot)
Tool Support (MCP)	Yes	Planned	Yes (built-in)	No (hermes internal tools)	Yes (always forwarded, not allowlisted)
Streaming	Yes	Planned	Yes	No	Yes
Best For	Heavy usage, tools	Large context, pay-per-use	Full Claude Code toolset	Multi-provider model access	Untrusted/isolation-sensitive agents

Using Copilot

workflow:
  runtime:
    provider: copilot
    default_model: gpt-5.5

Copilot is the default provider — runtime.provider can be omitted entirely. Requires an active GitHub Copilot subscription and the GitHub CLI authenticated (gh auth login).

Using Claude

workflow:
  runtime:
    provider: claude
    default_model: claude-sonnet-5

Set your API key: export ANTHROPIC_API_KEY=sk-ant-...

Using Claude Agent SDK (Experimental)

workflow:
  runtime:
    provider: claude-agent-sdk
    default_model: claude-sonnet-5

Requires the claude CLI to be installed and authenticated. Install the SDK: uv add 'claude-agent-sdk>=0.2.82'

Note: runtime.mcp_servers is supported — servers are translated into the SDK's own MCP config and attach alongside the built-in claude_code preset (a narrowing per-server tools: filter is refused, since the SDK cannot enforce one). Per-agent tool allowlists are not bridged: a workflow-level tools: block is rejected at conductor validate for any agent that omits tools: (it would otherwise inherit a list the CLI can't map). Omit tools: to grant the full claude_code preset; an agent's tools: [] disables the built-in tools, though declared MCP servers still attach.

Using Hermes (Experimental)

workflow:
  runtime:
    provider: hermes
    default_model: anthropic/claude-sonnet-5

Install the library: pip install hermes-agent

Using ACA (Experimental)

workflow:
  runtime:
    provider:
      name: aca
      pool_endpoint: "https://my-agent-pool.example.westus2.azurecontainerapps.io"
      inner_provider: copilot
    default_model: gpt-5-mini

The aca provider delegates an agent's entire agentic loop, tools, and MCP calls to a remote Azure Container Apps dynamic-sessions sandbox instead of running it on the host — useful for untrusted or isolation-sensitive agents (e.g. running arbitrary generated code). Unlike the other providers, aca requires the structured provider: form with a pool_endpoint pointing at an operator-provisioned ACA session pool (scripts/aca/provision-pool.sh) and azure-identity for authentication. Resolves its inner Copilot credential automatically via COPILOT_PROVIDER_BASE_URL → COPILOT_GITHUB_TOKEN/GH_TOKEN/GITHUB_TOKEN → gh auth token, so a gh-authenticated operator needs no ACA-specific setup. See examples/aca-coding-agent.yaml for a full end-to-end example.

See also: Claude Documentation | Hermes Documentation | ACA Documentation | Provider Comparison | Migration Guide

Using a Local / Custom LLM Endpoint (Ollama, vLLM, Azure OpenAI, ...)

runtime.provider also accepts a structured object that routes the Copilot SDK at any OpenAI-compatible / Azure / Anthropic-shaped endpoint. Useful for local inference (Ollama, vLLM, LM Studio) and managed deployments (Azure OpenAI):

workflow:
  runtime:
    provider:
      name: copilot
      type: openai                          # openai | azure | anthropic
      wire_api: completions                 # completions | responses
      base_url: http://localhost:11434/v1
      api_key: ${OPENAI_API_KEY:-ollama}
    default_model: llama3.1                 # match your endpoint's model name

The structured form is opt-in: a bare provider: copilot keeps the default GitHub Copilot routing. See examples/copilot-local-llm.yaml for the full example (including an Azure OpenAI variant) and Configuration Guide → Custom Provider Routing for environment-variable fallbacks, security notes, and validator rules.

CLI Reference

conductor run

Execute a workflow from a YAML file.

conductor run <workflow.yaml> [OPTIONS]

Option	Description
-i, --input NAME=VALUE	Workflow input (repeatable)
-m, --metadata KEY=VALUE	Workflow metadata (repeatable; surfaced in workflow_started)
--workspace-instructions	Auto-discover AGENTS.md / CLAUDE.md / .github/copilot-instructions.md and prepend to every agent prompt
--instructions PATH	Explicit instructions file (repeatable)
-p, --provider PROVIDER	Override provider
--dry-run	Preview execution plan
--skip-gates	Auto-select at human gates
--web	Start real-time web dashboard
--web-bg	Run in background, print dashboard URL, exit
--web-port PORT	Port for web dashboard (0 = auto)
-l, --log-file PATH	Write logs to file

Output verbosity is controlled by root-level options, which must appear before the subcommand:

conductor --quiet run workflow.yaml    # -q: minimal output (agent lifecycle and routing only)
conductor --silent run workflow.yaml   # -s: no progress output (JSON result only)

conductor validate

Validate a workflow file without executing.

conductor validate <workflow.yaml>

Full CLI documentation: docs/cli-reference.md

Workflow Registries

Conductor supports named workflow registries — GitHub repos or local directories containing shared workflows. Configure a registry once, then run workflows by short name.

Quick start

# Add a registry
conductor registry add official myorg/conductor-workflows --default

# List available workflows
conductor registry list official

# Run a workflow from the registry
conductor run qa-bot                       # latest from default registry
conductor run 'qa-bot@official#v1.2.3'     # specific tag (quote the #)
conductor run 'qa-bot@official#main'       # branch HEAD (re-resolved on fetch)

See docs/design/registry.md for the full design.

Examples

See the examples/ directory for complete workflows:

Example	Description
simple-qa.yaml	Basic single-agent Q&A
for-each-simple.yaml	Dynamic parallel processing
parallel-research.yaml	Static parallel execution
design-review.yaml	Human gate with loop pattern
script-step.yaml	Script step with exit_code routing
set-step.yaml	Set step deriving named values + boolean-routed branching
wait-step.yaml	Wait step + script for a polling loop-back pattern
wait-smoke.yaml	Minimal wait-only smoke test (no provider required)
terminate.yaml	Explicit type: terminate with success and failure paths

More examples and running instructions: examples/README.md

Documentation

Document	Description
Workflow Syntax	Complete YAML schema reference
CLI Reference	Full command-line documentation
Parallel Execution	Static parallel groups
Dynamic Parallel	For-each groups and array processing
Claude Provider	Claude setup and configuration
Hermes Provider	Hermes setup and configuration
ACA Provider	Azure Container Apps sandboxed execution setup and configuration
Provider Comparison	Copilot vs Claude vs Hermes decision guide

Development

Prerequisites

• Python 3.12+
• uv for dependency management

Setup

git clone https://github.com/microsoft/conductor.git
cd conductor
make dev

Windows

On Windows, use uv directly instead of make:

uv sync --all-extras    # instead of make dev
uv run pytest tests/    # instead of make test
uv run ruff check .     # instead of make lint
uv run ruff format .    # instead of make format

Copilot CLI path: Windows subprocess cannot resolve .bat/.ps1 wrappers by name alone. If you see [WinError 2] The system cannot find the file specified when running workflows, set the full path to the Copilot CLI:

# Find your copilot CLI
Get-Command copilot* | Format-Table Name, Source

# Set the path (use the .cmd variant from npm)
$env:COPILOT_CLI_PATH = "C:\Users\<you>\AppData\Roaming\npm\copilot.cmd"

Common Commands

make test             # Run tests
make test-cov         # Run tests with coverage
make lint             # Check linting
make format           # Auto-fix and format code
make typecheck        # Type check
make check            # Run all checks (lint + typecheck)
make validate-examples  # Validate all example workflows

Code Style

• Ruff for linting and formatting
• ty for type checking
• Google-style docstrings

Contributing

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit Contributor License Agreements.

When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.

To submit a pull request, follow these steps:

1. Fork the repository
2. Create a feature branch (git checkout -b feature/amazing-feature)
3. Make your changes
4. Run tests and checks (make test && make check)
5. Commit your changes (git commit -m 'Add amazing feature')
6. Push to the branch (git push origin feature/amazing-feature)
7. Open a Pull Request

Trademarks

This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.

License

MIT License - see LICENSE for details.