Changelog
What changed in Halto
We started building Halto in June 2026, and the first version went live in October. Below is what changed for customers, newest first.
ImprovedClearer pricing
Each plan shows its limits for your AI agents and your people's computers, with what every plan includes listed once, and the docs have a page for the browser extension. We also fixed some bugs.
ImprovedA clearer Overview, new pricing for teams, and fixes
The top of Overview shows every decision on one chart and what needs your look beside it. A Team plan for companies comes after the beta, and Halto stays free while it is in beta. We also fixed some bugs.
ImprovedImprovements and fixes
The dashboard is easier to use, and we fixed some bugs.
ImprovedThe dashboard on a phone
Tables keep the decision on screen, the Employees and agent buttons line up, tabs wrap instead of hiding, dates no longer split in two, and the menu's close button no longer covers the workspace switcher.
NewAn account menu, and help when something goes wrong
Your name at the bottom of the side menu opens a new account menu: settings, plan and this month's hidden values, docs, status, what's new, light or dark, language, currency, and Talk to us. Error messages carry a Tell us button that sends us the error. We reply to your email, usually within a day.
ImprovedSeven times more hidden values on the free plan
The data guard now hides up to 350,000 values a month on the free plan, up from 50,000: one table of customers read by Claude Code can hold hundreds.
ImprovedSwitch the currency from the account menu
Owners and admins can switch the workspace's currency (euro, dollar, pound) from the account menu. Amounts keep their numbers, so €100 becomes $100, and the switch asks for Face ID.
NewHalto for teams
A page on how Halto reaches every employee's computer: device management installs Halto Guard, each person's Claude Code and Codex work under their team's rules, and every decision is logged.
ImprovedSite visits, counted without cookies
We now count visits to this website to see where visitors give up: anonymously, with no cookie and nothing stored in your browser, never linked to an account, and not at all when your browser sends Global Privacy Control. The privacy policy lists what is noted.
FixedTemplates in your currency
A template's limits and their sentences follow the workspace's currency: a dollar workspace reads "Refunds over $50 need my approval". An amount in another currency than a rule's says why it counts as over the limit.
ImprovedDocs in pages, with search
The developer docs are now pages in four groups, with ⌘K search, code in TypeScript, Python and curl that stays in the language you picked, and a copy button on every block.
SecurityTenth security review
Our own review of everything new since the ninth: people and teams, company computers, the browser extension and the Slack app. Every confirmed finding is fixed and kept as a test; a new computer now waits for an admin where it could skip someone's rules.
NewDeveloper docs, a security page and this changelog
One docs page for every way to connect (SDKs, MCP gateway, Halto Guard, model gateway, REST), a security page that also says what Halto doesn't protect against, and a status page.
NewHalto in the browser, in testing
An extension for Chrome and Edge stops a message to ChatGPT, Claude, Gemini or Copilot that holds what the company forbids. It joins with the company's key, like Halto Guard.
NewEvery computer at once
Jamf, Kandji and Intune run one installer that locks Halto Guard into Claude Code and Codex, where the person using the computer can't turn it off. Each computer joins with a key of its own.
NewEmployees sign in to their own history
A person on the Employees page sees only what Halto stopped or hid for them and why. A company's proven domain adds new people by itself.
ImprovedBusy, not broken
Under a burst, agents get 503 database_busy with retry_after instead of an error, and the SDKs wait and retry. Dashboard pages update by themselves as agents work.
NewRules for people and teams
A rule can be for one person or one team as well as for agents. Blocks still come first, so nobody's own allow opens what a rule for everyone forbids. Requests go to the team's approvers.
NewApprove from Slack
Halto's own Slack app: approve or reject with buttons, and only a linked approver's press counts.
NewClaude's tokens count in budgets
Through the model gateway, each answer's cost is worked out from Anthropic's prices, and a spent budget stops the next request.
SecurityTwo more security reviews
Rounds eight and nine of our own review: every confirmed finding is fixed and kept as a test.
SecurityFail closed, always
When Halto can't be reached or can't decide, the action is blocked everywhere. The per-agent setting and the SDKs' failMode: "open" are gone (SDK 0.2.0).
NewData rules in plain language
Say where a kind of data may go: "Guest phone numbers may go only to cloudbeds.com", "Never send passport numbers outside the business".
NewThe signed log
Every few minutes each workspace's new decisions are sealed with Ed25519, each seal chained to the one before. Download it from Reports and check it at /verify-log.
NewPython SDK on PyPI
pip install halto: check, protect and guard, like the TypeScript SDK.
ImprovedFace ID approvals in the dashboard
A request that needs Face ID can be approved from the dashboard: the same button asks for Face ID, Touch ID or the phone, bound to that one request.
ImprovedHalto Guard on Windows, Linux and Alpine
One program per system, no Node.js needed, checked against its checksum when installed.
NewHalto Guard for Claude Code and Codex
Runs on the developer's computer: data is hidden before the agent reads it and every command is checked before it runs. No Claude login or key reaches Halto.
NewThe model gateway
Claude Code and the Claude API through Halto under your own key: placeholders go to Claude, real values come back to your computer, commands are checked first.
Improvedn8n node with provenance
n8n-nodes-halto is published from a public repository with npm provenance.
ImprovedData guard across countries
Personal data is found the way each country writes it: ID numbers, phone numbers, IBANs and addresses.
NewNames and addresses, and where data goes
Halto's own model, on Halto's server, can find people's names and street addresses too. A new page shows which kinds of data each agent sent and where.
NewTypeScript SDK on npm
npm install @halto/sdk: check an action, wait for an approval, report what happened.
NewThe data guard
Before an agent's model reads what a tool returned, Halto replaces emails, phone numbers, IBANs, card and ID numbers with placeholders. The real values go back only into actions your rules allow.
ImprovedEight more languages
The dashboard, emails and approval messages in German, French, Spanish, Italian, Portuguese, Dutch, Polish and Greek.
SecurityA second factor at sign-in
Once someone has Face ID or a fingerprint set up, every new sign-in asks for it, and owners and admins must set it up within 7 days. Google's sign-in screen now names Halto.
NewThe Monday digest and rule suggestions
A weekly email of what your agents did, and suggested rules for the requests you keep approving.
Newn8n node and approval webhooks
n8n-nodes-halto on npm, and webhooks that tell your own systems when an approval is decided.
SecurityAn outside review of the code
An architecture and code review from outside Halto. Every finding is fixed and kept as a test.
NewApprove from your phone
The Halto app on the phone's home screen: push notifications, approvals with Face ID, and an emergency stop for every agent.
MilestoneThe first version
The check API and the MCP gateway, rules written in plain language, approvals in Telegram and by email, the Attack lab and the oversight report.
MilestoneWork on Halto begins
The idea and the first designs. The code followed in October.