CorenelGet started

Product

One agent, six ways to work.

Corenel is a single agent runtime. Every surface below points the same policy-gated worker at a different kind of job: same guard rails, same recording, different shape of task.

Agent

A conversation that touches real files.

Ask for a change, watch it read its way to an answer, and review what it proposes before anything is written. Use it when the job needs judgement and you want to stay in the loop.

How to run an agent
you raise the retry limit to five
  read src/client/retry.ts
  gate write src/client/retry.ts

Crew

Agents that start themselves.

A crew member runs on a schedule, on a signal, or on its own judgement: a job that fires without a chat you have to remember to start. Run the sidecar as a daemon and it keeps firing whether or not the browser is open.

Each member is a folder of plain files: its instructions, a soul, a persona and its own policy. An orchestrator member can take a task, submit a plan for your approval, hand the pieces to the rest of the crew, run the checks you confirmed, and open a pull request, asking at each step, or not, by the autonomy level you choose.

How to build a crew
On a scheduleA cron expression, in UTC
On a signalWhen a run finishes or a file lands
Self-directedDecides for itself when there is work

Craft

Documents you watch take shape.

The agent writes and the document renders live beside it, so you are editing a result rather than reading a description of one. Export to PDF, HTML or DOCX when it is right.

How to craft a document
draft.html
  live preview
  export: pdf · html · docx

Workflows

The parts that must not drift, don’t.

Deterministic steps (API calls, mappers, branches, loops) sit beside agent steps in one flow. Put judgement where you need judgement and put a fixed step everywhere else.

How to author a workflow
fetch → map → agent → branch
                          ↳ notify

Sessions

A run can get its own machine.

Work happens in a session the daemon holds open. Turn on microVM isolation and each session is its own sandbox behind a kernel boundary: pause it, snapshot it, come back to it tomorrow, or destroy it when the job is done. The sidecar tells you at startup which mode it is in.

How sessions work
PauseStops where it is, keeps its state
SnapshotA point you can return to
DestroyThe sandbox and its contents go

Recordings

Scrub back through what happened.

Every agent and workflow run is captured: every tool call and every decision, in order, with a scrubber. Recent runs roll off as new ones arrive; pin the ones you want kept. It is how you answer what an agent did without taking its word for it.

Watch one now →
14:02:11 read src/client/retry.ts
14:02:15 gate write src/client/retry.ts
14:02:23 approved by you

Same worker. Same rules.

Whichever surface you start from, the policy, the sidecar and the recording are the same ones.

Free, and you bring your own model key.