Conductor

The coordination layer for an agentic organization. AI agents do the work — author specs, write code, run campaigns, maintain the wiki. Humans review, approve, and steer. Conductor is where the two meet: shared work items, automated workflows, a living knowledge base, and connections to the tools your organization already uses.

• Work Items & Reviews — typed work (PRDs, issues, campaigns, anything) with documents, line-level comments, reviewers, and approval gates. Types and statuses come from your workflows, not hardcoded enums.
• Workflows — two complementary kinds:
  • Lifecycle statecharts define how a Work Item moves through states (draft → review → approved → build), including which agent or human acts at each step.
  • YAML automation runs jobs on schedule, webhook, or event triggers — with agent, claude-code, docker, http, and integration action steps, artifacts, secrets, and loops. See docs/workflows.md.
• Agents — bring-your-own-key model personas with tools drawn from connectors and custom HTTP tools. Agent runs are fully transcripted for observability. Any agent can opt into conversations — multi-turn, addressable-by-name chat via the REST API or Discord's /ask, defaulting to a project's self-healing CEO coordinator. Conversations feed a workspace-scoped memory — durable facts and decisions extracted, consolidated nightly, and injected back into future turns. See docs/conversations.md and docs/memory.md.
• Knowledge Center — an ingestion inbox feeding an agent-maintained wiki. Librarian workflows organize incoming knowledge into reviewed, versioned pages. See docs/knowledge.md.
• Integrations — a unified connector framework (OAuth, webhooks, fetch, actions): GitHub, Discord, GCP, Apple Search Ads, Google Search Console, PostHog, RevenueCat, Meta, YouTube, TikTok, and more. See docs/integrations-adding-a-connector.md.
• Publishing — take a Work Item through review to something live on a platform. Per-post scheduling on a calendar, an approval bound to the exact media and destinations it was given for, and per-destination outcomes with the live link. Publishes through Facebook, Instagram, YouTube and TikTok, or through a human by hand when there is no integration. See docs/publishing.md.

flowchart LR
    Human["Team member<br/>(browser)"]
    Agent["AI agent<br/>(Claude Code)"]

    subgraph Local["Developer machine"]
        Tools["@cliangdev/conductor<br/>CLI + MCP server + sync daemon"]
    end

    subgraph Cloud["Conductor"]
        UI["conductor-frontend<br/>Next.js 16"]
        API["conductor-backend<br/>Spring Boot 4 / Java 21"]
        Engine["Workflow engine<br/>triggers · jobs · steps"]
        DB[("PostgreSQL")]
        GCS[("GCP Storage<br/>documents")]
    end

    subgraph Exec["Step execution"]
        CloudRun["GCP Cloud Run<br/>managed or BYO runtime"]
        Worker["conductor-worker<br/>self-hosted runner"]
        Runner["runner-image<br/>step containers"]
    end

    Ext["Third-party services<br/>GitHub · Discord · GCP · …"]

    Human --> UI --> API
    Agent -- "MCP tools" --> Tools -- "REST" --> API
    API --> DB
    API --> GCS
    API --> Engine
    Engine --> CloudRun --> Runner
    Engine --> Worker --> Runner
    API <-- "connectors + webhooks" --> Ext

Human flow: review and approve work in the web app — documents, comments, reviews, workflow runs, knowledge pages.

Agent flow: Claude Code (or any MCP client) uses the conductor MCP server to create and transition work items, write documents, dispatch workflows, search and update knowledge, and call integration tools — no copy-paste.

Execution flow: workflow jobs run as containers on GCP Cloud Run (Conductor-managed or your own via runs-on runtime targets) or on a self-hosted conductor-worker next to your own Docker daemon.

conductor/
├── conductor-backend/     # Spring Boot 4, Java 21 — REST API, workflow engine, connectors
├── conductor-frontend/    # Next.js 16, TypeScript, Tailwind, shadcn/ui
├── conductor-tools/       # @cliangdev/conductor — CLI + MCP server (single npm package)
├── conductor-worker/      # self-hosted job runner (Express + Docker)
├── runner-image/          # container images for workflow step execution
├── docs/                  # architecture and contributor guides (see below)
├── docker-compose.yml     # local dev stack (db, backend, frontend)
└── Makefile               # dev, seed, e2e, logs, cli-install, ...

Prerequisites: Docker + Docker Compose, Node.js 22+, Java 21 + Maven (backend work only).

make dev         # build + start db/backend/frontend, seed demo data
                 # → http://localhost:3000 ([email protected] / conductor)
make logs        # tail all service logs
make down        # stop everything

Tests:

cd conductor-backend && mvn test        # backend (Testcontainers)
cd conductor-frontend && npx vitest     # frontend unit
cd conductor-tools && npx vitest        # CLI + MCP
make e2e                                # Playwright against the running stack

CLI + MCP for agent use: see the conductor-tools README — conductor login, conductor init, conductor mcp.

Contributions are welcome — bug reports, docs, and code. See CONTRIBUTING.md for local setup, PR workflow, and coding conventions. All contributors agree to our Code of Conduct and sign a lightweight CLA on their first PR (a bot prompts you automatically).

Found a vulnerability? Please report it privately via GitHub Security Advisories — do not open a public issue. See SECURITY.md for scope and response expectations.

Conductor is licensed under the PolyForm Noncommercial License 1.0.0. You are free to use, modify, and distribute Conductor for personal and non-commercial purposes — including hobby projects, research, and use by charitable, educational, or government organizations.

Commercial use requires a separate license. If you'd like to use Conductor commercially, please open a GitHub Discussion or contact @cliangdev.