Start here

Three things, in this order.

The order is not editorial. Without a sidecar the agent has no files to work on, and without a key no model answers, so the third guide does nothing until the first two are done.

Give the agent your machine

Pair a sidecar to let the agent work on real files: it is confined to the directories you approve.

1. Install the @corenel/sidecar package once (it needs Node 22 or later). It ships separately because it is a server with its own release line.
2. Run it on the folder you want to open up: corenel-sidecar --root <your folder>. It prints an address and a pairing token.
3. Open the sidecar dialog from the rail indicator, paste what it printed, and Connect (the dialog also shows the exact command for this machine).in the app
4. Pick the working directory from the source selector at the top of the agent workspace.

Add your API key

Add a provider key under Settings -> Models, or a custom OpenAI-compatible endpoint for a local or third-party model.

1. Open Settings to add an API key.in the app
2. Pick a built-in provider and paste your key; it is stored encrypted with your account, and the gateway uses it for your runs.
3. Running a local model, or using another vendor? Add it as a custom OpenAI-compatible endpoint instead; the browser calls that directly and the key never reaches us.

Work with the agent

Paste a screenshot straight into the composer: it goes to the model with your message.

1. Pick a model in the header; the picker also lists any agent running on your own machine, under "On your machine".
2. Type in the composer and send (Cmd/Ctrl+Enter).
3. Attach context with the + button: upload a document, or paste a screenshot straight into the composer.

Open Agent →

By surface

One guide per place you can work.

Each of these teaches a destination in the app, and ends with a link into it.

Build a crew

A crew agent can run on its own clock: a sidecar started as a daemon fires scheduled runs without the browser open.

1. Open Crew from the navigation rail.
2. Create an agent: give it instructions, pick its tools, and set its policy.
3. Run it on demand, or add a trigger. Schedule, Continuous and Self-directed triggers need the sidecar started with --daemon, which then fires them without the browser open.

Open Crew →

Craft a document

Craft turns a chat into a formatted document you can export as PDF, HTML, or DOCX.

1. In the workspace, open a project from the Projects list in the rail, or start one with its + button.
2. Describe the document you want in the chat; the draft appears in the document panel beside it.
3. Iterate in chat, then export from the panel header (PDF, HTML, or DOCX).

Open /app?mode=craft →

Replay a past run

Every run is recorded: replay it under Activity with the timeline scrubber.

1. Open Activity from the navigation rail; the Runs view lists them.
2. Pick a recording on the left (saved or recent).
3. Scrub the timeline on the right to step through the run.

Open Activity →

Author a workflow

API, mapper, branch and loop nodes never call a model: that is what makes them cheap, auditable, and the same on every run.

1. Open Workflows from the navigation rail.
2. Add nodes from the palette: an API call, a mapper, a branch, a loop, or a prompt, agent or guardrail node where you want a model to judge.
3. Wire them in order and set each node's inputs, then run it and watch the result at each step.
4. This page also opens saved strategies for review; a strategy is planned, so edit its steps by asking the agent to replan rather than from the palette.

Open Workflows →

Watch the machines your work runs on

Activity → Live shows every machine your sidecar is holding open: pause, snapshot, or destroy one from there.

1. Pair a sidecar first; with none paired this view has nothing to show and says so.
2. Open Activity and pick Live to see what is running, its CPU and memory against their caps, and its output.
3. Pause, snapshot, or destroy a session from here; a kernel boundary confines it when one is in effect.

Open /app/activity?tab=live →

Going further

Panels, proxies and the parts underneath.

These teach a panel, an external agent, or something the app does while you are not looking, not a place to navigate to.

Use Claude Code through Corenel

The model picker can run Claude Code on your own machine: your key, your policy, gated tool calls.

1. Start the sidecar with an external agent: corenel-sidecar --root <folder> --acp claude-code --acp-policy read-only.
2. Open the model picker and choose Claude Code, under "On your machine".
3. Chat as usual. Its tool calls are gated by your policy, and its own modes appear in the mode selector once the session starts.

Watch the agent plan its work

The Tasks panel shows the agent's own plan: watch a multi-step request move through Active, Blocked, Queued, and Done.

1. Ask for something that takes several steps; the agent breaks it into a task list as it plans.
2. Open the Tasks panel to watch the list as a timeline or a status board: Active, Blocked, Queued, Done.

See what the model is actually being sent

The Context panel shows exactly what is and is not being sent to the model, including anything offloaded or compressed.

1. Open the Context panel to see what is currently sent and what has been left out.
2. Watch context-window usage as the conversation grows.
3. Offloaded or compressed content is marked; restore an offloaded item if the agent needs it back.

In depth

The whole of a topic.

Longer guides for the parts that take more than a few steps.

Getting started

Sign in, add a model, give the agent a folder, and find your way around.

Running the sidecar

Install, pair, share folders, and run crew agents unattended.

Building a crew

Agents with their own soul, persona, instructions, tools and triggers.

Tasks and the orchestrator

From a task to a pull request, at the autonomy you choose.

Approvals, policy and Home

Where it asks, what it may do alone, and what it records.