Antler

In development

Early access

For security and IT teams

See, control and prove what AI coding agents do on your company's laptops.

Antler is a small recorder on each laptop and a dashboard for your team. You see the commands agents run and the credentials they use, and your rules can ask for approval or block an action before it runs.

In development. Not open for sign-ups yet.

For security and IT leads at companies whose engineers work with coding agents such as Claude Code, Codex CLI or Gemini CLI.

Antler's dashboard, the live log of the made-up company Acme: 528 records added today, and a list of what AI agents did on each computer, newest first. The newest action, a force-push by Claude Code, needs approval.
The dashboard's live log. Example data of a made-up company.

How it works

Three steps. No change to how engineers work.

  1. Install on laptops

    IT rolls a small recorder out through the device management you already use. One person can set it up with one command.

  2. Agents work, Antler records

    Each action a supported agent takes is recorded: who, what, which credentials and why. Your rules can ask for approval or block an action before it runs.

  3. Review, get alerts, show auditors

    Your team reviews in the dashboard and gets alerts in the tools it has. Auditors get a record they can check: it shows if anything was edited or removed.

Why Antler

What Antler stands on. Four things you can check for yourself.

  1. A record you can check without relying on us

    The record's formats are published, with test files, so anyone can write their own checker; Antler's verifier program checks an exported record offline. The check needs the export and the company's key fingerprints, and asks nothing of Antler's servers. Read the format What a passing check shows

  2. Coverage that shows its gaps

    For each agent and operating system, Antler says whether it is recorded before the action, after it, or only detected. What the recorder did not see is counted and shown. Which agents

  3. Start on your own, free for one person

    The plan is that one person or a small team can start alone: free for one person, with up to 3 devices, 30 days of history and no card. Sign-up is not open yet. Early access

  4. Built so that engineers can accept it

    Each person can see the records Antler holds about them, and a personal account's records keep only the facts by default. What leaves the laptop is listed, and the recorder adds about 4 ms per action, measured on Linux. What leaves the laptop

For security teams

Know what agents did. Decide what they may do.

Coding agents run commands, read files and use credentials on your engineers' laptops. Antler gives you the record of it, and rules that act before an action runs.

  • See each recorded action

    Who ran it, on which laptop, the command or file, the credentials used and what led to it.

  • Rules that ask or block

    Each kind of action gets an outcome: log, flag, ask or block. Built-in rules start watch-only, so you see what a rule would do first. A rule can name the program, the operation and the resource of common commands: git, the cloud command lines, kubectl, terraform. That reading is a hint: a command Antler does not understand matches no such rule.

  • Approvals in Slack

    When a rule asks, the action waits. Your team can decide approvals in a Slack channel. A rule's question can go to a group of approvers, such as a security team, instead of the developer.

  • Alerts in the tools you have

    A flag raises an alert in Slack, Microsoft Teams, a signed webhook, a syslog receiver or your SIEM, with a link to the dashboard. Alerts can also reach the person on call through PagerDuty or Opsgenie, or a Google Chat space. Those three follow each service's published API and were tested against local stand-ins, not live accounts.

  • A record that shows tampering

    Each record is signed on the laptop and linked to the one before. Edit or remove one and the check shows where. Try it

  • Credential recall, with a report

    When a key leaks, recall it and keep a signed report of what touched it. A recall does not revoke a key at its provider. How a recall works

  • What each AI tool can reach

    The AI inventory lists each AI tool's tool servers (MCP), its plugins and the types of credential it can read. Antler reads names and types from the tool's settings, never a value. What is kept outside those settings is not found.

  • Evidence you hold

    A workspace can have its log timestamped by an outside authority it chooses, and its records delivered to storage the company owns. antler-verify checks both offline. So far tested against a stand-in authority and a stand-in bucket, not real ones.

  • Legal holds

    An admin can hold the records of a person, a computer or a time range, with a name and a reason. Held records are skipped by retention and deletion, and each hold is itself recorded. Antler makes no legal claim about holds or timestamps.

Antler's dashboard, one record opened: Claude Code ran an upload to a storage bucket on a computer, with what led to it listed step by step, starting from the person's own prompt.
One recorded action, with what led to it. Example data.
Antler's dashboard, the Rules page: what a rule in watch mode does, and the watch-mode rules that matched in the last 7 days with what each would have done. One would have asked 13 times, another 5 times.
Rules, and what the watch-mode ones would have done. Example data.
Antler's dashboard, an approval request named Force push: the rule that asked, the computer, its owner, the kind of account and the permission mode.
An approval request, with who and what it is about. Example data.
Antler's dashboard, the Alerts page: what each kind of alert means, above a list that starts with a high alert, a bulk export from one computer. On a wide window a second alert follows: a new tool server came into reach of an AI tool.
Alerts, each with where it happened. Example data.
Antler's dashboard, the Credentials page: the secret itself is never stored, and the counts of what agents touched start with 1 credential, 3 touches and 1 sent to a host.
Credentials that recorded actions touched. Example data; values are never stored.
Antler's dashboard, a credential recall: a GitHub token pasted into an agent's prompt is recalled. The recall is active, 7 of 8 computers acknowledged it, the ones still unproven are named, and there is a button for the signed report.
A credential recall and the computers it reached. Example data.
Antler's dashboard, the AI inventory's list of what each tool can reach: Claude Code's tool servers by name, starting with figma and firebase, with how each is reached, over HTTP or as a local program, and what it runs or talks to.
What each AI tool can reach: names and types, never a value. Example data.
Antler's dashboard, the page Evidence you hold: outside timestamps are switched on, from the stand-in authority timestamps.example. There are 323 timestamps, and the newest covers 8,323 records.
Outside timestamps on the log, from a stand-in authority. Example data.

For IT

Roll it out with the tools you have. Know what it does on a laptop.

Antler is pushed like any other managed software. You can see what runs on a laptop, what leaves it, and how to take it off again.

  • Rollout through device management

    The dashboard's Deploy page renders the files for your workspace: Jamf Pro, Kandji, Microsoft Intune, other Windows tools, Linux.

  • Single sign-on and SCIM

    People sign in through your identity provider. With SCIM, removing a person ends their sessions and API tokens at once.

  • What runs on the laptop

    A small recorder that starts when an agent acts, writes the record and exits. If Antler's server cannot be reached, the agent keeps working and the records wait on the laptop.

  • What leaves the laptop

    A record of each action, with known secrets turned into tags first. Not your repositories. What leaves, in detail

  • Light on the laptop

    About 4 ms per action, measured on Linux. Memory: 11.6 MB at its peak. On GitHub's shared test machines: about 10 ms on a Mac, about 17 ms on Windows.

  • Uninstall

    An uninstall script comes with the rollout files. It removes the recorder and the managed settings Antler wrote.

Antler's dashboard, the Deploy page: a choice of operating system and device management (Jamf Pro, Kandji and Microsoft Intune for macOS, Microsoft Intune and PowerShell for Windows, Ansible for Linux), above the files for the one chosen, Jamf Pro.
The Deploy page: the files for a rollout. Example data.
On the laptop
One small recorder, and Antler's hooks in the agent's managed settings.
Per action
About 4 ms at the median, on Linux.
Leaves the laptop
Records of actions. Not repositories.
Network
HTTPS only, to your workspace in the US or the EU.
Systems
Built for macOS, Linux and Windows.
How this was measured

Median 3.89 ms and peak memory 11.6 MB with a typical policy of 80 rules, measured on 16 November 2026 on a 4-vCPU 2.1 GHz Intel Xeon virtual machine running Linux 6.18, with nothing else building or testing on it. In that run every one of the 19 hook cases had a median under its budget; the highest, the case for Codex CLI with cause links from its session file, was 4.38 ms. Mac and Windows were measured on GitHub's shared test machines on 17 November 2026 and are held to no budget. On an Apple silicon Mac the median was 9.82 ms and peak memory 16.2 MB. On Windows, with 2 CPUs and Microsoft Defender on, the median was 16.98 ms and peak memory 12.8 MB.

Rollout

How rollout works. Pushed by IT, or one command for one person.

Nobody installs Antler through the agents: there is no plugin or extension to add to an agent. It is set up on the laptop, by IT or by its user.

  1. Choose your device management

    On the dashboard's Deploy page, pick the operating system and the tool you manage it with.

  2. Download the files

    A profile, a configuration and scripts, rendered for your workspace.

  3. Push them like any other software

    Each laptop enrolls by itself and starts recording. Nobody on the laptop has to do anything.

For companies, a rollout through device management is built and tested, with packages for macOS, Windows and Linux. None is published yet.

For one person: one command, then antler login
  1. Install

    By install script, Homebrew or npm, when Antler opens. Nothing is published yet.

  2. Run antler login

    It opens your browser, where you sign in and approve the laptop. Your workspace is in the US or the EU.

  3. Records appear

    Login sets up Antler's hooks in Claude Code. From the next action on, records appear in your workspace.

antler login, as it runs
$ antler login --server https://us.app.antler.ai --no-browser
To connect this computer to Antler, open

    https://us.app.antler.ai/activate

and enter the code

    VLGZ-SMDB

Only approve it if you just ran this command on this computer. Waiting for approval (Ctrl-C to stop)...
Connected this computer to 'Alice Admin's workspace' (ws_7wij5urcjq2imi6mz44g) as device d_76ee7882c55b17d2286fe996454ac971.
Dashboard: https://us.app.antler.ai/w/ws_7wij5urcjq2imi6mz44g/
Applied the workspace's policy, version 3.
Hooks:   Claude Code (/home/alice/.claude/settings.json)
Binary:  /home/alice/.antler/bin/antler
Timer:   loaded (systemd antler-flush.timer, every 60 s)

What antler login printed in a run against Antler's development stack, with the stack's address replaced by the US region's. Alice is the stack's made-up person. The service is not open yet.

Coverage

Which agents, and what Antler cannot see.

On Linux, macOS and Windows: recorded before each action, and rules can ask or block. Checked in real runs. Claude Code is covered in the editor as well as the terminal.

  • Claude Code
  • Codex CLI

On Linux and macOS: recorded before each action, and rules can ask or block. Checked in real runs. On Windows that is not switched on, because a real run there did not pass: there Gemini CLI is recorded after the fact, as a log only.

  • Gemini CLI
  • GitHub Copilot CLI

Built, not switched on: no real run has checked them yet. Found on a device, not recorded.

  • Cursor
  • Windsurf
  • Cline
  • Factory Droid

Built and tested, not open to anyone yet: inviting teammates, company single sign-on, device-management packages. Still planned: blocking in Gemini CLI and GitHub Copilot CLI on Windows.

Antler's dashboard, the Coverage page: 99%, one square for each session of the last 7 days. 76 of 77 checked sessions were fully recorded, 1 partly, and 9 are unknown and not in the percentage.
The Coverage page counts what was recorded and what was not. Example data.

Not sampled. Every action the hook sees is recorded; what it did not see is counted. Antler compares its records with Claude Code's own transcripts and shows each session as complete, partial or unknown. A missed action is a gap you can see.

  • 45 tool calls in Claude Code's transcripts that week
  • 44 recorded
  • 1 gap

For Claude Code on a company account, the signed coverage report also compares the vendor's own record with the laptops' records, session by session: recorded fully, partly or not at all. Sessions that reach the model another way, such as through an API key, and sessions of the other agents are listed as not comparable and are never counted as matched.

Not seen today: commands a person types themselves, agents on machines without Antler, and programs that read a credential on their own. If a developer switches Claude Code's hooks off, Antler finds the session in Claude Code's own transcripts and shows it as a gap, unless those were switched off too.

Sign in with GitHub, Google or Microsoft.

Built for macOS, Linux and Windows.

Example

One session, as Antler records it. Tamper with it and see.

Larkfield is a made-up company. One engineer asks Claude Code to fix a failing test: eight actions, each on the record. Change one and the check shows where.

The record UTC mira-mbp Claude Code

15 signed records

Each record is one line, signed by the laptop it was made on. Each line carries a fingerprint of the line before, so the lines are linked. Change or delete one and the check fails, and shows where.

Records Time Tool What Outcome Fingerprint
Records 1 to 8 to A read, a failing test, an edit, a passing test. logged
Records 1 and 2 Read internal/ledger/round.go logged 86f67438
Records 3 and 4 Bash go test ./internal/ledger/... logged exit 1 5b5479fe
Records 5 and 6 Edit internal/ledger/round.go logged 79e2fae9
Records 7 and 8 Bash go test ./internal/ledger/... logged exit 0 8859f765
Records 9 and 10 WebFetch https://forum.devtips.example/t/go-rounding-half-even/4127 logged outside content 1b1e12e3
Records 11 and 12 Bash STRIPE_API_KEY=[REDACTED:stripe_key:40841d54] go test -tags integration ./internal/charge/... logged key tagged 4bf9a9ac
Record 13 Read ~/.aws/credentials Blocked by Larkfield's rule a0288288
Records 14 and 15 Bash git push origin fix/rounding Flagged by Larkfield's rule a built-in would have asked 0e254730

One square, one signed record: 8 actions, 15 records. Each picture is drawn from its record's fingerprint.

One Claude Code session at Larkfield, a made-up company. Example data.

UTC mira-mbp Claude Code

See the whole session, step by step What a passing check shows, and what it does not

Questions

Short answers. What people ask first.

Does it slow my agent down?

Not that you would notice. The recorder adds about four thousandths of a second to each action (about 4 ms at the median), against actions that take seconds. That is measured on Linux. On GitHub's shared test machines it was about 10 ms on a Mac and about 17 ms on Windows. The check happens before the action so that a rule can ask or block before something runs, which reading what happened afterwards cannot do. A workspace that wants no delay at all can have an agent recorded after the fact instead; rules then cannot ask or block for that agent.

What happens if Antler breaks?

By default the action runs, and the fault is recorded and raises an alert. A workspace can choose to fail closed instead.

Can a user switch the hook off?

Yes. A person can remove a hook from their own laptop, and Antler does not pretend otherwise. When a laptop's records stop or the links between them break, an alert is raised, and the coverage meter counts tool calls that ran without a record. In a company rollout the hooks sit in managed settings, Antler reports when another settings source overrides them, and a built-in rule blocks an agent that tries to edit Antler's own files.

How is this different from a network gateway's AI security?

A gateway watches the network and the model conversation: which AI services a device reaches, and what an agent tells a model when routed through it. Antler records what the agent does on the machine: commands, files, local tools, credentials used. With Claude Code, before the action runs. Records are signed and gaps are counted. Antler works alongside the gateway you already have.

What about personal Claude accounts?

In a company workspace, a personal account's records keep only the facts by default: no prompt, command, path or output. Keeping content needs a notice to employees.

What leaves my laptop?

A record of each action, with known secrets turned into tags first. Not your repositories. What leaves the laptop

Is Antler certified or audited yet?

Not yet. No SOC 2 Type II or ISO/IEC 27001 audit has started yet: both are planned and being prepared. An independent penetration test is planned before the company launch and is not yet commissioned. Compliance and testing

Is Antler open source?

No. You can still check the proof yourself: this page checks it in your browser, and antler verify checks a laptop's own copy.

Integrations

Fits the tools you already run.

Coding agents
Claude Code, Codex CLI, Gemini CLI, GitHub Copilot CLI. How far each is covered
Sign-in and directory
GitHub, Google or Microsoft accounts, single sign-on through your identity provider, SCIM.
Approvals and alerts
Slack, Microsoft Teams, Google Chat, PagerDuty, Opsgenie, Splunk, Microsoft Sentinel, syslog, signed webhooks.
Device management
Jamf Pro, Kandji, Microsoft Intune, other Windows tools, Linux.
Evidence
A download for auditors, checked offline with antler verify, and signed search exports. Outside timestamps, delivery to storage you own, and legal holds.

Built and tested here, most of them against a local stand-in and not yet with a customer's real account.

See each integration and its status

Where Antler fits

On the machine. Beside your gateway and your endpoint security.

A gateway sees what crosses the network. A tool server sees what reaches it. Antler records what the agent does on the machine, including what never leaves it.

On the machine
Commands, file reads and writes, local tool servers, the credentials used: Antler.
On the network
Which AI services, uploads, remote tool servers: your gateway.
At the tool
Who may call what: your identity and access tools.
On the endpoint
Malware, processes and the health of the device: your endpoint security tool.

Antler works beside them. Its records and alerts can be sent to the log and alerting tools your team already has.

Security and privacy

Each workspace kept apart. In the region you choose.

Kept apart
Each workspace has its own database and its own keys: no table holding your records, devices, rules or keys is shared with another workspace. The keys are sealed with a cloud key service.
In transit
Laptops and browsers reach Antler only over HTTPS.
At rest
Disks are encrypted. Each workspace's backups are encrypted with a key of its own.
On the laptop
Secrets in the formats Antler knows become tags before a record is written or sent.
Device keys
Each laptop signs its records with a key of its own. A laptop can keep a second key in its security chip (a TPM on Linux, the platform provider on Windows), so that a copy of its files on another computer shows. This is off by default and not offered on macOS. It has run against stand-ins, not yet a physical chip, and that the key is in a chip is the computer's own statement.
Region
Your workspace's data stays in the region you choose, the US or the EU. Sign-in through GitHub, Google or Microsoft, and Slack if you use it, are US services.

Compliance and testing

  • SOC 2 Type II Preparing The audit has not started yet: the control documents are drafted, and no auditor is engaged.

  • ISO/IEC 27001 Preparing No certification body is engaged yet, and no audit has started.

  • Independent penetration test Planned Planned before the company launch, and not yet commissioned.

Nothing is certified or audited yet. When audit and test reports exist, they will be shared with customers under NDA.

Early access

Free for individuals. Paid plans for companies.

For one person
Up to 3 devices, 30 days of history, no card.
For companies
Paid plans.

Antler is in development and not open yet.

Sign-ups Not open

Sign-ups are not open yet. They open once the privacy notice is published.