Hatchet vs Orchestrator Zero

Yes

Workers register with Hatchet from any environment, local machine or VM.

docs.hatchet.run

Conditional: Mac and Linux; Windows via WSL

Workers are your own Python, TypeScript, Go or Ruby processes; the quickstart that sets one up covers macOS, Linux and WSL, and native Windows is not documented.

docs.hatchet.run

Yes

The engine is open source under MIT and can be self-hosted.

docs.hatchet.run

Yes

Workers open long-lived gRPC connections to the engine, so nothing calls into them.

docs.hatchet.run

Conditional: Labels, no version rollouts

Worker labels route work to matching workers, but there is no versioned rollout of workers.

docs.hatchet.run docs.hatchet.run

Yes

A durable task replays from its last checkpoint after a worker crash.

docs.hatchet.run

Not documented

Hatchet exposes tasks as tools and provides durable execution. Its docs do not describe a platform policy that allows, refuses or sends each tool call for approval.

docs.hatchet.run

Conditional: Event waits; chat is your code

A durable task can wait for an event pushed through the API or SDK, which is the documented basis for human approvals; there is no built-in approval step or chat app integration, so a Slack button is a webhook you wire up yourself.

docs.hatchet.run hatchet.run

Not documented

Searched the docs for tool definition changes, rug pull, tool pinning and tool poisoning; Hatchet exposes tasks as tools to agent SDKs and does not document recording or checking tool definitions.

Yes

Each tenant is a separate environment with its own workflows and workers.

docs.hatchet.run

Not documented

Nothing found on spend limits, budgets or cost caps in the docs or on the pricing page.

No: Keys in worker environment

The agent guides set the provider key as an environment variable in the process that runs the agent; Hatchet has no model gateway.

docs.hatchet.run

Not documented

The Claude Agent SDK guide describes a trusted-process setup, not an isolation boundary. A built-in sandbox with a network host allowlist is not documented.

docs.hatchet.run

Conditional: Paid plan, not tamper-evident

Paid plans expose audit logs for actions such as API tokens, membership and API-triggered runs. A hash chain or signature for verifying agent activity is not documented.

docs.hatchet.run hatchet.run

Not documented

Searched the docs for GDPR deletion, per-user keys and crypto-shredding; the middleware guide shows encrypting payloads with a key held on the worker, but nothing per person.

Conditional: Retention, no legal hold

Retention of runs and events is set per tenant on a self-hosted engine and by plan on Hatchet Cloud, and nothing documents a legal hold.

docs.hatchet.run

Conditional: Your code calls the model

Hatchet does not call models itself; its agent guides use the Claude Agent SDK and the OpenAI Agents SDK, and any other vendor or a local model is code you write in the worker.

docs.hatchet.run

Conditional: MCP, no A2A

Tasks and workflows can be exposed as MCP tools and the CLI includes an MCP server; A2A does not appear in the docs.

docs.hatchet.run docs.hatchet.run