Conductor CLI
A Microsoft CLI that runs multi-agent workflows from one YAML file: no model in routing, parallel steps, web dashboard.
- 499 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- Script steps run shell commands and agents work with project files and outside services.
- Installation uses a network script that installs `uv` and downloads a release.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add cli/microsoft-conductorDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Script steps run shell commands and agents work with project files and outside services.
- Installation uses a network script that installs `uv` and downloads a release.
This is third-party code. Review the repository files before installing.
What it does
Conductor describes multi-agent workflows in one YAML file: agents, their prompts and the routing between them. Routing uses Jinja2 templates and expressions and the first matching condition wins, so no model spends tokens choosing the next step. The same input follows the same path, and workflow files can be diffed in pull requests and run the same way locally and in CI. Run one with `conductor run workflow.yaml --input question="What is Python?"`. Providers are GitHub Copilot and Anthropic Claude, plus experimental Claude Agent SDK, NousResearch Hermes and Azure Container Apps sandboxed execution. Steps can be parallel or dynamic, and there are sub-workflows, script steps, computed-value steps and MCP tool calls without a model, plus a pause for a human decision. The `reasoning.effort` setting translates to each provider native options. The `--web` flag opens a dashboard with the workflow graph, and Fleet Manager lists all running processes.
Who it is for. Teams building pipelines from several agents, such as code review, research with synthesis or plan then implement, who want repeatable results.
Good fit when
- Code review runs as a chain of several agents and the step order must match on every run.
- The workflow must live in the repository and changes should be reviewed in pull requests.
- Agents must run in parallel over a list and the results get merged.
- An agent must stop and ask a human before the next step.
Not a fit when
- One agent in a regular chat is enough and workflows are not needed.
- Orchestration is wanted where the model itself decides what to do next.
- There is no access to GitHub Copilot or an Anthropic key.
Example request
Describe a three-agent workflow in YAML, with a researcher, a critic and an editor, and run it through `conductor run` with a question about Python.Limitations
Access to GitHub Copilot or Anthropic Claude is required. Claude Agent SDK, Hermes and Azure Container Apps execution are marked experimental in the README. Installation uses the `https://aka.ms/conductor/install.sh` script, which installs `uv` when needed. Updating uses the same script from a new shell: in-process self-upgrade of a running Conductor is not supported. Workflows with script steps run shell commands, so read YAML from other repositories before running it. The last commit in the repository is from November 2026.
How to disable. Delete the Conductor virtual environment that the install script created and its directory with the `conductor` executable.
FAQ
How do I install Conductor?
With `curl -sSfL https://aka.ms/conductor/install.sh | sh` on macOS and Linux or `irm https://aka.ms/conductor/install.ps1 | iex` on Windows. The script verifies a SHA-256 checksum.
Does orchestration spend tokens?
No, routing uses templates and conditions with no language model. Only the agents themselves spend tokens.
Which providers are supported?
GitHub Copilot and Anthropic Claude as main ones. Claude Agent SDK, NousResearch Hermes and the Azure Container Apps sandbox are marked experimental.
Can I watch a workflow in a browser?
Yes, `conductor run workflow.yaml --web` opens a dashboard with the graph, an event stream and prompts for human decisions. `--web-bg` runs it in the background.
Which steps exist besides agents?
Scripts routed on exit code, computed-value steps, MCP tool calls without a model, sub-workflows and an explicit terminate step with a reason.
Discussion
No questions or reviews yet.
Related
CLI toolsA CLI and plugin: agent workflows live in the repository as Markdown and YAML with roles, scripts, pauses and branching.
A local Node.js orchestrator that runs Codex, Claude Code, Gemini and OpenCode as one team led by an operator agent
Skills and the porch CLI for agent orchestration: a strong model plans, cheaper ones code, several model families review.
One agent launches others in the background: Claude Code, Codex, Copilot CLI, Grok Build and Pi, with a model chosen per stage.
CLI Agent Orchestrator
CLI Agent Orchestrator (CAO)
A CLI that coordinates several coding agents: a supervisor delegates tasks to specialists in parallel or in sequence across isolated tmux sessions
An orchestrator of an AI expert team: from one sentence it assembles roles that run YAML workflows to a finished result with acceptance checks
README in short
Deterministic agent orchestration: conditions compute the routes with no language model, and the workflow lives in the repository as plain YAML.