Orchestra 0.9.0

Orchestra

Deterministic AI agent orchestrations. Describe a workflow as a declarative DAG of steps in a single JSON/YAML file, and Orchestra runs it — resolving dependencies, streaming progress, and driving each Prompt step on a real coding agent (GitHub Copilot or OpenCode). It ships as one self-contained command-line tool with a built-in web portal, a one-shot runner, MCP integration, triggers, checkpointing, and human-in-the-loop pauses.

inputs ─▶ ┌─ research ─┐                       orchestra init
          │            ├─▶ brief ─▶ document   orchestra doctor
          └─ (Copilot) ─┘         (no model)   orchestra run hello

Highlights

• Zero to running in two commands. orchestra init scaffolds a working workspace; orchestra doctor verifies your machine before the first run instead of failing mid-run.
• One file, one DAG. Steps (Prompt, Command, Script, Http, Transform, Approval, nested Orchestration) wired by dependsOn, with template expressions ({{param.x}}, {{step.output}}).
• Pluggable agents. Run any Prompt step on copilot or opencode, selectable per step or per orchestration.
• MCP-native. Attach Model Context Protocol servers to steps, and expose your orchestrations as an MCP server.
• Agent Skills. Drop SKILL.md directories onto a step so the agent gains specialized, on-demand workflows.
• Operate it. A web portal, triggers (cron/webhook/file/manual), durable checkpoint/resume, profiles & tags, and a full CLI — all under the single orchestra command.

Install

Orchestra is distributed as the Orchestra command-line tool (command: orchestra) on NuGet. Your orchestrations are just JSON/YAML — no C# required.

# Run without installing (recommended for trying it out) — requires the `dnx` launcher:
dnx Orchestra --yes -- <command> [options]

# …or install it as a global tool:
dotnet tool install --global Orchestra
orchestra <command> [options]

Running orchestra with no command prints the help. orchestra <command> --help documents any subcommand.

Quick start

orchestra init             # scaffold a workspace with a runnable example
orchestra doctor           # verify prerequisites before the first run
orchestra run hello        # run it
orchestra new my-workflow  # then add your own, from a template

No agent set up yet? orchestra init --template smoke-test --yes && orchestra run smoke-test gives you a green run in about a second with nothing configured - it proves the install before credentials enter the picture.

init writes a starter orchestration under orchestrations/, the JSON schemas under .orchestra/schemas/ for editor autocomplete, and an orchestra.json whose scan block is what lets you run orchestrations by name. It prompts for the template, provider, and model — pass any of those as flags to skip the prompt, or --yes to accept every default:

orchestra init ./my-workflows --template research --provider copilot --yes

Five templates ship with the tool, and every one runs with no arguments:

doctor checks the things that otherwise only fail during a run: which orchestra.json is in effect and whether it parses, whether the data path is writable, whether the agent CLI is present and authenticated, and whether a configured server is reachable. The first Copilot run downloads the Copilot CLI (~100 MB, once per machine) — orchestra doctor --fix does it up front instead of mid-run. If credentials are missing, orchestra login runs the provider's own sign-in flow.

Orchestra discovers orchestra.json by walking up from your working directory, so a scaffolded folder is self-contained; a user-global config at %APPDATA%\Orchestra\ (or ~/.config/Orchestra/) applies everywhere else.

validate parses an orchestration and checks its expressions without a server, an agent, or any cost — it exits 0/1/2 so CI and Script steps can branch on it:

orchestra validate ./orchestrations/hello.yaml

Your first orchestration

orchestra init writes this for you as orchestrations/hello.yaml; here it is in full. It's a DAG: research produces findings, brief summarizes them, and document shapes the result without calling a model at all.

# yaml-language-server: $schema=../.orchestra/schemas/orchestration.schema.json
name: hello
description: Research a topic, brief it, then assemble a document.
defaultModel: claude-opus-4.8

inputs:
  topic:
    type: string
    required: false
    default: deterministic AI agent orchestration

steps:
  - name: research
    type: Prompt
    systemPrompt: You are a meticulous research assistant.
    userPrompt: Research "{{param.topic}}" and list the key findings as concise bullet points.

  - name: brief
    type: Prompt
    dependsOn: [research]
    systemPrompt: You are a technical writer.
    userPrompt: |
      Using the research below, write a 120-word executive brief on {{param.topic}}.

      {{research.output}}

  - name: document
    type: Transform
    dependsOn: [brief]
    template: |
      # {{param.topic}}

      {{brief.output}}

Because topic has a default, orchestra run hello works with no arguments — and --param topic="vector databases" points it somewhere else.

Run it

orchestra run runs one orchestration to completion. With dnx (no install):

dnx Orchestra --yes -- run --run-file ./orchestrations/hello.yaml --param topic="vector databases" --report markdown

--mode decides where it runs:

• auto (default) — attach to a running Orchestra instance when one is configured (via --server, the ORCHESTRA_URL env var, or the hostBaseUrl/urls in your discovered orchestra.json) and healthy; otherwise spawn a throwaway in-process host just for this run.
• existing — require a healthy configured instance (error if none).
• isolated — always run self-contained. orchestra exec is shorthand for run --mode isolated.

orchestra run hello --param topic="vector databases"   # registered, on your server (auto)
orchestra exec --run-file ./orchestrations/hello.yaml  # self-contained one-shot
orchestra portal                                       # long-running host + web UI

The portal (orchestra portal) serves the dashboard, REST API, and MCP endpoints; the client verbs (orchestra list, get, register, attach, runs, triggers, profiles, tags, pending, respond, …) talk to it over HTTP/SSE.

Step types

Agent providers

Every Prompt step runs on an agent provider, chosen by precedence step provider → orchestration defaultProvider → host default:

defaultProvider: opencode
steps:
  - name: draft   # runs on OpenCode (the orchestration default)
    type: Prompt
    userPrompt: "..."
  - name: review
    type: Prompt
    provider: copilot   # this step overrides to Copilot
    userPrompt: "..."

• copilot — GitHub Copilot CLI via the Copilot SDK (supports the full feature surface).
• opencode — spawns an opencode serve HTTP server; supports MCPs, subagents, reasoning, working dir, skills, engine tools, attachments, permissions, excluded tools, infinite-session toggle, and worker swap with session resume.

The engine fails a step fast if it uses a feature the chosen provider doesn't support (rather than silently dropping it). See docs/copilot.md and the provider notes for the full capability matrix.

Agent Skills

Agent Skills are SKILL.md directories that give a Prompt step specialized, discover-on-demand knowledge and workflows. Reference them per step with skillDirectories (relative paths resolve from the orchestration file):

steps:
  - name: code-review
    type: Prompt
    skillDirectories:
      - ./skills/code-review
      - ./skills/security-analysis
    userPrompt: "Review these changes: {{param.changes}}"

A SKILL.md is Markdown with a small YAML frontmatter (name, description) plus the instructions and any reference files. Orchestra ships an orchestration-authoring skill (under skills/) that teaches an agent to write valid Orchestra files. It ships inside the tool too, so orchestra init --with-skill drops a copy into .orchestra/skills/ without cloning this repo — that's what the generate template uses to write orchestrations from a plain-English description:

orchestra init --template generate
orchestra run generate --param description="review my open PRs every morning and post a digest"

The generated file is machine-checked with orchestra validate before it's written, so a draft that wouldn't parse fails the run instead of landing on disk.

MCP integration

Attach Model Context Protocol servers (local stdio or remote HTTP) to a step so the agent gains tools:

mcps:
  - name: zakira-recall
    type: local
    command: dnx
    arguments: ["Zakira.Recall", "--yes", "--", "mcp"]
steps:
  - name: research
    type: Prompt
    mcps: [zakira-recall]
    userPrompt: "Use the web tools to research {{param.topic}} and summarize."

Orchestra also exposes your orchestrations as an MCP server (/mcp/data), so other agents can discover and invoke them. See docs/host.md.

Documentation

License

MIT — see LICENSE. Repository: https://github.com/MoaidHathot/Orchestra.