Security
Controls for teams that ship.
The code stays on your machines, review is a gate you configure, and membership, models and permissions are visible instead of implicit.
The gate
Bot review, with escalation when it matters.
With review on, a finished task goes to a reviewer that checks it against its acceptance criteria. It can approve, or send back a list of concrete changes. The builder never approves its own work.
Bot review is a project setting. Turn it off and finished tasks wait for a person instead.
- A reviewer, not an editorThe reviewer is instructed not to edit files; it returns a verdict and concrete notes.
- Three rounds, then a personIf the reviewer and builder do not converge, the task stops in Needs you with the notes.
- Failures do not hideA failed reviewer run gets three tries in total, then a person is asked.
Landing is a setting, not a prompt. Leave the branch, open a PR, merge, or merge and push. Failed landings show up in Needs you.
- Round 1 · changes requestedAdd alt text to the illustration.
- Round 2 · approvedAcceptance criteria met.
- LandingLeave branchOpen PRMergeMerge + push#42 Redesign the 404 pageOpened
The boundary
The hub coordinates. Your repositories stay home.
The hub stores and coordinates projects, scopes, tasks, reviews, conversations, usage, devices and org settings. The code, branches and worktrees live on the machines you connect. The hub never needs a copy of your repository.
- Repositories and worktrees (connector clones under ~/.factory; a source can be any git folder you choose)
- Agent processes and their sessions
- Connector logs
- Projects, scopes and tasks
- Reviews, threads and messages
- Usage rows, artifacts and attachments
- Members, roles and skills
- Devices, versions and heartbeats
- Permission rules and notifications
Coding agents call their own model providers; the hub does not sit in that path.
The org
Org-wide controls, project-level when you need them.
Membership, roles and model defaults apply to the organization. Permission rules and auto-approval are set per project.
Four access levels
Owner, admin, member and viewer. Admins manage members, roles, channels and devices; only an owner can grant owner, and the last owner cannot be removed.
Roles are teammates
A role carries a prompt, skills and a default agent and model. The planner assigns one to each task, you can change it, and each role keeps its own conversation per scope.
Defaults that stick
Choose a default model per agent for the organization, then override it per project, role or task.
Permission rules
Approve once to remember the exact request, or give an agent auto-approval in a project. Rules are listed and can be removed; auto-approval is shown with a YOLO warning. Members can edit these in project settings.
The ledger
See the work. Get told when it matters.
Token usage by project, model, scope and task. Connect Telegram or Slack and choose which events reach you.
Token usage
Demo App · all time| Model | Input | Output | Thinking | Cached | Total | Runs |
|---|---|---|---|---|---|---|
| claude-sonnet | 820k | 96k | 41k | 310k | 1.3M | 14 |
| codex | 240k | 22k | — | 110k | 372k | 6 |
Usage is counted in tokens. Some agents also show an estimated cost per thread. Agents that do not report usage are left out.
Pick smaller models for roles that do not need the big one.
Alerts
- Task ready for reviewon
- Task failedon
- Agent needs permissionon
- Agent has a questionon
More events available.
Needs you also flags a running task with no activity for 20 minutes.
Still have a governance question?
The FAQ covers data, review, roles, usage and stalls.
Factory Fox