The Agents app (named Neural Inverse Agents on your system) is a desktop app for delegating work to AI agents. You start sessions, run several in parallel, each in its own worktree if you like, and review the changes each agent makes before they reach your branch. It is the agent-first window of Code - OSS, with Neural Inverse sign-in, models, governance, audit, context and team work.
It is installed together with the Editor and runs on its own, with its own process, icon, data folder (.neural-inverse-agents), link scheme (neuralinverse-agents) and sign-in session.
Maturity
The Agents app is pre-release. Interfaces and data formats may change between versions. Sharing running agent sessions between the Agents app and the Editor is in progress, and the agent host shared by both apps is on by default only on macOS. Features that talk to your organisation's Govern deployment have, in places, only been tested against servers written from the published contract. Each page says what is verified.
The window
| Area | What it shows |
|---|---|
| Sessions list | Your sessions, grouped by workspace or date, on the left. It can be collapsed. |
| Chat | The conversation with the agent of the selected session. Several sessions can be shown side by side. |
| Side pane | Tabs for Changes (every file the session changed, as diffs), Files and Browser, with a detail panel for the branch's changes or the file explorer. |
| Title bar | Your account, and the connection to your organisation's Govern deployment: signed out, connecting, connected (Keystone: <region>), degraded with a reason code, or locked by policy. The Agents window has no status bar. |
What you can do
| Task | Page |
|---|---|
| Run several sessions, organise and archive them, compare models | Sessions and parallel work |
| Start a task, choose a folder, worktree, agent and model, review and apply the result | Starting and reviewing a task |
| Open a pull request from a session, start a session from a pull request | Pull requests |
| Understand what your organisation's policy decides, approve and sign | Approvals and governance |
| Schedule agent sessions (experimental) | Automations |
| Move between the Agents app and the Editor, and know what they share | Working with the Editor |
| Look up settings | Settings |
| Fix a problem | Troubleshooting |
Get started
- Sign in to your organisation. The Agents app signs in as its own OAuth client,
neuralinverse-agents, so its sessions, its entry in your device list and its audit records are its own. Sign-in works as in the Editor: see First launch and signing in. - Connect a model on the Models page (Manage Models, in the Manage menu). The Agents app shares the Editor's model connections and keys. See Models.
- Choose New Session, pick a folder, describe the task, and submit. See Starting and reviewing a task.
How the Agents app is governed
- Every agent action is decided before it runs. The agent host decides each tool call, shell command (after full parsing), file read and write, host contacted, MCP server and model, for every engine. The answer is allow, deny, or needs approval; in a session nobody attends, needs approval is a refusal.
- Engines. Sessions run on the engines of the agent host: Claude (Claude Agent SDK), Codex (only on a ChatGPT or OpenAI Codex account, off by default), and your connected models through the agent host's bring-your-own-model bridge. Governance always wins over an engine's own settings: Claude's bypass mode is turned off, Codex's approval policy and sandbox are set from the rules, and a rule an engine cannot enforce is reported, never dropped.
- Built-in protection. Even with no organisation policy, an agent cannot change policy and approval records, git hooks, CI configuration, shell profiles or audit and governance data without a person.
- Audit. Every session, prompt and tool call is recorded in the device's activity record, shared with the Editor. Nothing leaves the device unless your organisation turns on uploading to its own Govern deployment.
- What binds. A modified build, or an engine run outside the agent host, is outside the device's controls. What must hold is enforced by Govern (its model gateway, organisation context, vault) and, for merges and releases, by your CI.
See Approvals and governance and Organisation policy.
Running from source
./scripts/agents.sh # macOS / Linux
.\scripts\agents.bat # WindowsThe Agents app uses its own Electron bundle (.build/electron-agents) and data folder; its identity overrides are in product.agents.json.