Skip to content

FAQ

Short answers, each linking to the page that covers it properly. If you are stuck on a specific symptom, common issues is indexed that way instead.

CommandChain is in a closed beta. Apply from the product page, and you will get an invitation by email once your application is reviewed. See the quickstart for what happens after that.

You set direction, and a team of AI agents shapes, builds, launches, and grows the work, while you keep the decisions that matter. See what is CommandChain.

What does an agent do without me watching?

Section titled “What does an agent do without me watching?”

Exactly as much as you have allowed. Every class of action sits on an autonomy ladder from “ask me each time” through “act, then show me” to “act within a policy”, and some classes are locked so they always need a person. See approvals and the glossary.

No. Coding work is one of the things the platform does, not the way you use it.

Onto a machine you choose. A runner checks the repository out, runs the coding agent against it, commits, and pushes. On a self-hosted runner that machine is yours, and your source never leaves it. See workspaces and git.

Your workspace (meetings, tasks, runs, projects, and the audit log) lives in the hosted control plane. Your source code does not: it is checked out on the runner that does the work, and on a self-hosted runner that machine is yours. See workspaces and git.

No. Every agent process runs inside an operating-system boundary the runner builds before it starts, with a synthetic home directory holding only the credential for that one attempt. If a machine cannot enforce that boundary, the runner refuses to run an agent there at all. See the runner security model.

It is stored encrypted for your workspace and released to a single run attempt at a time. A second request for the same release is refused. That is a security property, not an error. See agent credentials.

Whichever you allow. A workspace sets a default model and an allowlist, and each agent can be pinned to something specific. See the model catalog.

No. A trial workspace is funded by trial credits, and after that the default is that agents think through a coding agent on a runner you pair, using a login that machine already has. A provider key is needed only for Direct-LLM mode. See connect an AI credential.

The request stops before it calls a model and returns a 402 with the error key trial_credits_exhausted. Nothing is charged and no partial work happens. Pair a runner or connect your own credential to keep going. See trial credits.

For real work, yes. Anything that touches a repository, a filesystem, or a command line happens on a runner. Conversation-only workspaces can run on Direct-LLM without one.

What does the installer put on my machine?

Section titled “What does the installer put on my machine?”

One executable in ~/.local/bin, a 0700 directory at ~/.commandchain, a user-level service, and a log directory. No administrator rights, no runtime, no package manager, and no token to paste. See install a runner.

Not a self-hosted one. There is no unprivileged Windows sandbox this project has proven, and a runner that cannot contain an agent is not one we will install. Windows uses a managed runner. See runners overview.

Check spam first. If it is not there, the invitation record still exists: whoever sent it can copy the link from the invite dialog and send it to you directly, or re-send. See users and invitations.

You can receive notifications and use slash commands in Slack. Approvals happen in the web app.