Run there.
Work here.
Give agents remote compute. Keep your work close.
We’re building a way to work with agents on your machine, a remote host, or your team’s cluster—without losing track of the task, its permissions, or its results.
See the experience ↗Your machine isn’t
the limit.
The agent may need more compute, or access to an environment your laptop doesn’t have. You should still be able to follow the work from where you are.
The goal is one way to interact with work, wherever it runs. The connected local-to-remote experience is still in development.
Follow the task.
Not the session.
A worker can stop. A connection can drop. That shouldn’t erase what the task is, what has already happened, or the results you’ve kept.
Habitat gives the work its own identity, separate from the agent carrying it out. We call that an Operation. Its recorded history and retained results outlive an individual worker.
This gives recovery something to build on. It doesn’t mean every interrupted task can resume exactly where it stopped.
Give it a task.
Give it limits.
An agent shouldn’t need the keys to everything just to finish one job. Habitat ties limited permissions to the work and checks which worker is currently responsible for it.
A replacement worker doesn’t get to claim authority just because it says it owns the task. And delegating work must not silently expand what an agent is allowed to do.
These controls are runtime foundations today. Full credential brokerage and broader production integrations are still being developed.
Less reconnecting.
More getting on with it.
One task to return to—not a trail of sessions to reconstruct. This is the proposed workflow, not an available end-to-end experience.
- Choose where it runs.
Connect to your machine or an authorized remote environment.
- Set the work and its limits.
Give the task an identity, a worker, and only the permissions it needs.
- Follow progress. Come back later.
Inspect the same task rather than depend on one connection staying open.
- Pick up what’s been kept.
Retrieve results and review the record before deciding how to continue.
Clear answers.
No magic promises.
01What happens if the worker fails?
The work keeps its identity, journal, and retained results. Those records help establish what happened and what can safely happen next.
They are not a snapshot of every open file or the worker’s memory. Moving an unfinished workspace between hosts is still in development; automatic recovery of every task is not promised.
02What if an agent changes another system?
Habitat supports selected HTTP actions through a controlled runtime path. It does not mediate all network traffic or guarantee that every external change can be undone.
A record that says “request sent” is not proof the change succeeded. An uncertain result needs checking before retrying—not a blind repeat that could do the same thing twice.
03Can an old worker keep acting?
Habitat tracks who currently holds responsibility for the work. Custody checks and generations let the runtime reject a worker acting under an outdated custody state.
That does not instantly revoke credentials or short-lived proofs already issued to another system. Those systems still enforce their own rules. Runtime authority also does not replace process, container, or VM isolation.
04Can we trust the records?
Records should make actions easier to examine, not ask you to accept an agent’s story. A signature can establish who signed a record and whether it changed; it cannot, by itself, prove that every claim is true.
Independent verification also needs the relevant authority, supporting evidence, and its limits. Gaps and uncertain outcomes must stay visible.
05Where does the idea come from?
Plan 9 explored resource namespaces: the working environment need not be defined by one machine. MOSIX explored separating execution from location.
Habitat draws on those ideas for contemporary agent work. It is not a hypervisor, a Plan 9 compatibility layer, or a promise to move running processes transparently between machines.
Read the engineering background ↗Built foundations.
More to connect.
Habitat is in active development. The runtime foundations exist; the complete remote working experience does not yet.
- Implemented foundations
- Work identity, custody and generations, narrowing grants, journaling, retained results, and supported mediated HTTP effects.
- In development
- The connected local-to-remote workflow, full credential brokerage, and cross-host handoff.
Implemented foundations are not a claim of general-purpose production readiness.
How we describe availability
- Available
- Implemented and verified for the stated workflow and environment.
- Experimental
- A running prototype with known limitations.
- In development
- Planned or partially implemented capabilities.
What would you run
beyond your laptop?
Bring a real task. Let’s look at what it needs to run, recover, and stay under your control.