Code review
Rex reads a PR diff and its relevant callers, then returns severity-ranked findings with file locations and evidence.
macOS beta · local-firstLocal-first
Local-first
engineering digital employees
Configure agents, workspaces and memory on your Mac for code review, defect repair and technical debt inspection. Write actions pass through an approval gate and run in a Git worktree sandbox.
Agent execution, workspaces, memory and the database stay on your machine. Prompts and context required to complete a task are sent to the LLM provider you choose.
Verified capabilities
Let agents do work, while keeping the work inspectable.
This launch page lists workflows and boundaries that are verified today, rather than turning a roadmap into a promise.
Defect repair
Fixer takes a review finding, edits an isolated worktree, runs checks and delivers a diff.
Debt inspection
Janitor scans repositories read-only and organizes follow-up work; scheduled autonomy is off by default.
Multi-agent orchestration
Chain review, repair and delivery into a workflow, while each employee keeps its own role, skills and memory.
Approval and sandboxing
File writes and side effects pass through an approval gate; write-capable runs use a Git worktree.
Tools and connectors
GitHub PR and Feishu loops are verified; MCP, Slack and TypeX connectors depend on configuration and version.
Data boundary
Execution stays local. Model requests leave through your chosen provider.
AI Agent Console is not a central control plane that uploads your workspace. The platform runs on your Mac and connects directly to the LLM provider you configure when a task needs inference.
- Stored locally
- Agent execution and workspaceLocal files, Git repositories, run records and SQLite data are managed by your Mac.
- Visible memory
- File-based and editableMemory can be inspected, changed and removed instead of living in an invisible central account.
- Model requests
- Direct to your providerPrompts, code excerpts or context required for a task go to the provider you select.
- External connectors
- Enabled by your configurationConnectors and channels participate only when you authorize and trigger their workflow.
Your Mac
Agent loop, tool execution, code repositories and inspectable work records.
- workspace
- memory
- SQLite
- worktree
- approval
- audit log
Your chosen LLM provider
Receives the prompts and context needed for the current task. Provider terms govern its data handling.
- LLM request
- provider API
This is an architecture description, not a claim that no data ever leaves your machine. Choose providers, connectors and workspaces that fit your requirements.
One real loop
From finding a problem to delivering a reviewable change.
Permissions and deliverables follow the run, while a person keeps the decision at the critical write boundary.
Give Rex a PR
Read the diff and investigate context; read-only work does not change the source repository.
Confirm the defect
Return severity, file location, evidence and a suggested fix as a traceable review record.
Hand it to Fixer
Edit and check inside a Git worktree sandbox to produce an isolated diff.
Approve and deliver
Writes, merges and external side effects require your approval; deliverables remain local.
Current launch scope
Start with the part you can trust today.
The macOS beta is for engineering teams comfortable configuring Python, an LLM provider and a workspace. Packaging and signing are still being completed.
macOSCurrent launch platform, with a local setup script and console.
BYOKUse your own provider and API key; the provider charges model usage.
Human gateWrite actions and external side effects keep an approval boundary.
Put one employee on your repository first.
The macOS beta currently requires Python 3 and an internet connection to install dependencies. The signed and notarized download will be provided later; for now, use the install notes and contact entry to request beta access.