Dorf

Durable AI workers on infrastructure you control.

Dorf makes your coding agent a durable worker—Codex today, more harnesses over time. Give it a job, walk away, and come back to the same work.

Each Worker gets a full development environment for its native tools and workflows. The Room keeps that work isolated from the rest of your infrastructure.

Run a coding Job

Install Dorf and run the one-time guided setup:

uv tool install dorf
dorf setup

See Getting started for current prerequisites and setup details.

Then create a Worker and give it one complete goal:

dorf worker spawn python-developer

dorf job assign build-retry-queue \
  --to python-developer \
  --goal "Build a small Python retry queue with tests, run them, and publish the results as evidence."

dorf job wait build-retry-queue

Close the terminal while the Job runs, then return from another shell:

dorf job inspect build-retry-queue
dorf job message build-retry-queue "Add a test for exponential backoff and publish the new results."
dorf job wait build-retry-queue

Workers do Jobs in Rooms

Resource	Meaning
Worker	Durable identity around an agent harness
Job	Pinned goal with its own conversation and retained evidence
Room	Isolated environment supplied by a Room provider

An Assignment records the exact Worker, Job, and Room binding for lifecycle and recovery. Most human-facing commands need only the Worker or Job name.

Workers wrap the harnesses developers already use; they are not a replacement agent framework. The control plane is harness- and sandbox-provider agnostic by design. Codex and local Incus VMs are the only supported adapters today; other harnesses and local or cloud VM providers can fit the same boundaries once validated.

Why Dorf

• Leave without losing the thread. Worker and Job identity survive the client that started them.
• Return to the situation, not a transcript. Inspect state, timeline, and evidence without asking the Worker to recap.
• Let the harness do its best work. Give it a complete development environment and contain that environment with a clear isolation boundary.
• Build workflows above the runtime. Dorf owns durable execution; applications own policy, acceptance, and presentation.

Alpha status

Dorf is alpha software. See Setup support for current adapters, host requirements, and limitations.

Coding-to-PR showcase

Coding-to-PR is the first application built on Dorf, not the runtime's identity. It takes a coding task from an isolated workspace through checks, review, follow-up, and a PR proposal.

dorf start "Fix the checkout timeout and add a regression test"
# Or leave an issue running unattended:
dorf afk 42
dorf job inspect JOB
dorf job inspect JOB --json  # structured outcome pulse for agents
dorf verify JOB
dorf publish JOB

New coding Jobs use the Provider Connection selected by dorf setup. Pass --provider-connection NAME only to intentionally override that host-local default.

The official dorf-codex image is the coding workstation: repository preparation and checks come from the cloned repository's contract, while GitHub access remains a separately scoped host-side integration. See the North Star for how it fits the broader runtime.

Build on Dorf

Build workflow-specific applications with the experimental Python SDK. It exposes the same Worker, Room, Job, and Assignment controls used by the CLI; coding-to-PR is the first example.

from dorf import Dorf


def run_coding_workflow(task: str) -> str | None:
    with Dorf.open(provider_connection="personal-chatgpt") as dorf:
        dorf.spawn_worker("developer")
        dorf.assign_job(
            "implementation",
            worker_name="developer",
            goal=(
                f"{task}\n\n"
                "Acceptance:\n"
                "- Run the repository checks.\n"
                "- Publish the results as evidence."
            ),
        )
        result = dorf.wait_for_job_input("implementation")
        if result.outcome != "done":
            raise RuntimeError(result.detail or result.outcome)
        return result.response

The SDK is experimental and does not yet carry a third-party compatibility promise. See the Runtime Surface for the current responsibility boundary.

Learn more

• Documentation — choose a path by what you are trying to do
• Getting started — create your first Worker and Job
• North Star — approved product direction
• Setup support — host support, diagnostics, and troubleshooting
• Contributing — development workflow and DCO sign-off
• Security policy — private vulnerability reporting

Dorf is licensed under the Apache License 2.0.