
conductor
Runs workflows written as JSON on a server you host, resuming them after a crash — infrastructure, not a library.
What is conductor?
Conductor runs workflows you write as JSON — a list of tasks wired together by reference name — on a server you host, and it saves the result of every task, so a run that dies resumes at the task it stopped on instead of from the top. The code that does the real work sits outside the engine in workers: ordinary programs that poll the server over HTTP for one type of task and report the result, so any language that can call an HTTP API can host one and nothing in them has to be written a special way. Recent versions pull LLM calls, MCP tool discovery and vector search into the workflow as built-in task types, which makes an agent loop resumable without any framework code. The price is operational: this is a JVM server with a database, a distributed lock and an index behind it, and the open-source build ships no authentication of its own.
What can you do with conductor?
- Describe the workflow as JSON, then start it from anywhere — A definition is a list of tasks, each with a reference name that later tasks read outputs from. A run begins from the REST or gRPC API, from an event arriving on Kafka, NATS, RabbitMQ or SQS, or from the built-in cron scheduler.
- Workers are ordinary programs that poll for work — Your task code sits outside the engine: it polls the server over HTTP for one type of task, does the work, and reports success or failure. The engine never replays it, so clocks, random values and any library you like are fine. Java, Python, Go, JavaScript and C# have settled SDKs; the Ruby and Rust ones are marked incubating.
- A run survives a crash, and can wait weeks for a person — Every task result is persisted, so a restarted server picks up where it stopped. A HUMAN task holds the workflow open until someone marks it complete through the task update API, with no timer or waiting state of your own to keep.
- Branch, loop and fan out with built-in operators — SWITCH picks a path and DO_WHILE repeats a block until a condition says stop. FORK_JOIN runs branches in parallel, FORK_JOIN_DYNAMIC decides how many branches to create while the workflow is already running, and the JOIN that follows collects what they returned.
- Call an LLM, an MCP server or a vector index without writing a worker — LLM_CHAT_COMPLETE reaches Anthropic, OpenAI, Gemini, Bedrock, Ollama and others, LIST_MCP_TOOLS and CALL_MCP_TOOL work against any MCP server over HTTP, and the indexing and search tasks cover Pinecone, pgvector and MongoDB Atlas. Enabling a provider is a server-side job — set its API key as an environment variable and restart — after which a workflow picks between the enabled ones by parameter.
- Replay a finished run, and leave old versions running — The UI shows every task's input, output, timing and retry history, and you can restart a run from the beginning, rerun it from one task, or retry only the step that failed. Definitions are numbered versions, and a run stays on the definition it started with unless you deliberately restart it on the latest.
Before you choose conductor
- Nothing in the open-source server authenticates callers: the REST and gRPC APIs and UI are open to whoever reaches them, so it needs a proxy in front, and role-based access control is part of Orkes' commercial build.
- It is a JVM server you operate rather than a library you import, and the deployment guide's simplest production stack is still one PostgreSQL for storage, queueing and search plus a Redis for the distributed lock.
Star history
17 Aug to 30 Oct · +305
Frequently asked questions
Is conductor free for commercial use?
conductor is released under the Apache-2.0 licence — OSI-approved open source, which permits commercial use.
How can conductor be deployed?
conductor is available as Self-hosted / Managed cloud.