A terminal built for agent work.
Anteirus Terminal is a macOS terminal for working with command-line coding agents. Each agent’s session reads in full, from its reasoning to every tool call, diff and test, and a footer under each pane shows where it stands: its context, changes and plan, and, with hooks configured in the agent, its cost and whether it needs you.
In developmentUsed internally at Anteirus since mid-2026. Not yet released, so there is no download.
Walkthrough: an agent adds a data-leak check to a reward function
Interface previewScripted walkthrough with sample data. It runs no commands and sends nothing anywhere.
The session view, the footer and the pane states are modelled on the current internal build. The “needs you” state relies on hooks configured in the agent. The agent, the project and every number shown are samples.
Step 1 of 6 · Task. You give a coding agent a bounded task in a sample environment: fail any episode whose observations expose an unmasked policy number, then run the evaluation suite.
In this walkthrough, answer for the agent. In a real terminal you would type your reply.
The interactive steps need JavaScript. The same walkthrough is described in text below.
01 — Sessions
Every step an agent takes, readable.
What working with an agent looks like in Anteirus Terminal, and the real status of each part.
The whole conversation as history
In current build
In an agent’s pane, history becomes the agent’s conversation: its reasoning, each tool call with its result, diffs, and long output folded. It is read from the agent’s own session log on your Mac. This works with Claude Code (with a hook that announces its log), Codex CLI, Factory Droid and Grok. By our code review, it reads local files only and opens no network connections.
A footer under every agent pane
In current build
Who and what, context use, work in progress, plan and mode, in five rows. Model, cost and account fields need hooks configured in the agent; missing data shows as a dash.
See which session needs you
Prototype
Agent panes show a state, and the window and its tabs show the counts. The app works out running and done by itself; needs you, error and stalled come from hooks configured in the agent or from an internal status tool. Only needs you draws attention with motion.
Answer without switching panes
Prototype
A small card on a waiting pane sends a one-key yes or no, or a one-line reply. Each reply is checked before it is sent and logged on your Mac.
Approve or decline
Agent feature
Agents ask for approval through their own permission prompts. Anteirus Terminal shows the prompt; it does not enforce an agent’s permissions. What an agent may do is set in the agent and in your system.
Clean history for full-screen programs
In current build
Full-screen redraws, from tmux or an agent’s own interface, no longer leave stacked copies of the screen in history.
Review changes in your own tools
Changes an agent makes are files in your working tree, so you review and commit them with git and your editor as usual.
03 — Fleet
Many agents, one view.
The fleet console shows every session as a card: what it is doing, whether it is waiting for you, and a reply box. Approve, decline or message a session without finding its pane.
mask-observations3/4
mask policy numbers where observations are built
approve: edit envs/claims/observe.py
Update(envs/claims/observe.py)
feat/mask-at-source +2
message session
rotate-fixtures2/5
regenerate fixtures after the schema change
approve: run fixtures --regenerate
Bash(fixtures --regenerate)
chore/fixtures-v4 +118
message session
ledger-sandbox1/3
block outbound calls from the ledger sandbox
exit 1 · network policy test failed
Bash(make sandbox-test)
fix/egress-policy +14
message session
reward-review4/7
review reward checks changed this week
Read(envs/support/reward.py)
chore/reward-review
message session
episode-replay5/6
replay failed episodes with the new mask
Bash(replay --failed --env claims)
feat/replay-failed +41
message session
retention-sweep6/6
expire dataset snapshots past retention
report written
chore/retention +9
message session
Interface previewPrototypeSample data. Automatic re-organisation is off by default. F5 opens the console from any app, so the app takes that key system-wide.
04 — Recovery
Back to work after a crash.
Sessions run in a separate server process, so a session can be re-attached if the app quits unexpectedly. In our own testing this has re-attached single sessions; re-attaching many panes at once has failed, so treat it as best effort.
Today 09:41–11:06Crashedclaims-env· 3 panes · 14 min agoYesterday 16:20–18:02Closedledger-recon· 2 panes · 18 h agoOct 22 10:05–12:40Restoredsupport-routing· 4 panes · 2 days ago
claims-env · 3 panes in 2 folders · last active 14 min ago
plannerenvs/claimsPausedPlan 3/4coderenvs/claimsLostPlan 2/7reviewerevals/claimsPausedNo plansnap-3b7e, taken 4 min before the crash: conversations and pane layoutInterface previewPrototype · internalSample data. The recovery center reopens work by having each agent resume its saved conversation in a new process. It shows what it will reopen and changes nothing until you confirm. Running processes do not survive a restart of the Mac, and work in flight is redone. These tools are internal and not part of anything we distribute.
05 — Evaluating it
What exists today, and what does not.
If you are considering Anteirus Terminal for a team, this is its honest state. We would rather tell you now than in a procurement call.
Where each capability lives
| Capability | Where it lives | Status |
|---|---|---|
| An agent’s conversation as history | The app, reading the agent’s own session files on your Mac | In current build |
| Pane footer | The app. Model, cost and account fields need hooks configured in the agent. | In current build |
| Attention states (needs you, error, stalled) | The app, fed by hooks configured in the agent or by an internal status tool | Prototype |
| Quick replies and the fleet console | The app | Prototype |
| Optional session helper | A helper bundled with the app; sessions do not start in it by default | Prototype |
| Session recovery center | A separate local service with its own panel in the app | Prototype · internal |
| A record of which agent ran where | A separate local service | Internal |
For an engineering team
| Requirement | Today |
|---|---|
| Existing workflows | Works with your shells, editors, version control and command-line tools. Agents are ordinary programs in a session. |
| Deployment and updates | No installer, managed deployment or Anteirus update channel yet. Internal builds are installed by hand. They are signed with a local certificate rather than an Apple Developer ID, and are not notarised. |
| Platforms | Apple-silicon Macs only. The current build declares macOS 12.4 or later as its minimum; we develop and use it on macOS 26. |
| Model providers | Use the agent tools your organisation already uses, with your own accounts and under their terms. Anteirus does not resell model access or route agent traffic through its own servers. |
| Security review | A security policy with a named responsible person and a reporting process. No third-party certifications, and no independent security assessment yet. |
| Licensing | Licence terms will be published with the first release. Commercial terms for teams have not been set. |
| Support and commercial terms | No published pricing, support plans or service levels. |
| Administration | Not available: central administration, single sign-on, role-based access, managed deployment, audit export. |
06 — The walkthrough in text
The same session, step by step.
It uses a sample environment. The agent, the project and its output are illustrative.
Give the agent a bounded task
You ask a coding agent to make a reward function fail any episode that exposes an unmasked policy number, then run the evaluation suite.
It investigates
It searches the environment and reads the code that builds observations and scores episodes. Each tool call and its result appear in the session.
It changes the reward
Two added lines, shown as a diff in the session. The footer counts the change.
It runs the evaluation
Two sample episodes now fail, which is the point: both leaked a policy number into an observation.
It asks before going further
Fixing the source would change what every environment sees, so the agent asks first, and its pane switches to “needs you”.
You decide
Say yes, and the agent masks the field where observations are built and re-runs the suite. Say no, and nothing else changes. Reviewing and committing the work stays with you, in your own tools.