An owner per agent1
Every agent has a person or team accountable for it. No owner in the policy, nothing runs.
dono: Operações · RevOps papel_do_dono: operador
HARNESS · GOVERNED CLOUD IDE · EARLY ACCESS
The cloud IDE where every agent has an owner.
Skills from Agents & Skills, synthetic data from Evidence, policy per data class, human approval on sensitive actions and a verifiable receipt for every run.
Cloud IDE and runtime in early access, by invitation. The demo runs in your browser and executes nothing.
Send a follow-up to stalled contacts
leituraallowedescrita_arquivoapproved · simulatedenvio_externowaiting for approval
receipt #39c41e2a7b0f3…no action performed
WORKS WITH
Skills in the open format SKILL.md, the same one used by Claude Code and Codex
Agents & Skills open and free catalog
Evidence synthetic data for rehearsal
ON ONE SCREEN
What it does, what it does not do and how to get in.
The limits sit next to what works. Whoever signs off on the purchase needs to see both together.
Does today
- Opens one shell per person in the browser, with their own Claude Code or Codex account.
- Opens one session per branch, with its own worktree and agent; an agent on main can orchestrate the others.
- Protects main in the policy and in git itself: an agent only changes a protected branch with your approval.
- Runs every Claude Code tool through the policy before it executes: it proceeds, waits for a person or is denied.
- Records decision, approval and result in a signed receipt that anyone can check with the public key.
- Installs skills from Agents & Skills into the workspace with one command.
Does not do
- It does not govern Codex: Codex has no equivalent hook, and its actions produce no receipt. The git guard on main applies to it.
- It does not stop the workspace owner from running commands outside the agent, inside their own container.
- It does not yet limit internet destinations or disk quota per workspace.
- It is not yet a place for client data or personal data: the external security review comes first.
How to get in
- Join the waitlist and tell us the case you want to run.
- Access opens by invitation, one workspace at a time.
- The invitation works once, for 7 days, and only for you.
- No published price. The AI model cost is on your own account.
THE PROBLEM
An agent without an owner is risk without an address.
Teams already run agents across different tools. Nobody knows who answers for each one, what data it touches and who approved its last action.
- Agents without ownersnobody answers when something goes wrong
- Sensitive data in informal usecustomer bases pasted into any window
- External actions without approvalemail, CRM and ERP changed without a person seeing it
- Invisible costAI spend scattered across teams and cards
- Decisions without a trailno record of who asked, what ran and why
HOW HARNESS ANSWERS
Four guarantees before any action.
Fail closed: if a check fails, the request is denied and the reason is recorded.
Policy per data class2
Public, internal, confidential or personal. Each class requires a scope, designed with the LGPD in mind.
classes_de_dado: clientes.csv: pessoal
Human approval3
Sensitive actions wait for a person, and the requester cannot approve their own request.
aprovacao_obrigatoria: - envio_externo
Verifiable receipt4
Every step becomes a receipt chained by SHA-256. Changing a field breaks the chain.
#3 envio_externo aprovado
9c41e2a7b0f3…What each guarantee covers
- In the runtime console, every request carries the token of a workspace with an owner. In the Cloud IDE, the owner is the invited person, with a workspace of their own. ↩
- In the Cloud IDE, the default policy asks for a person when the input carries personal or confidential data. Classification reads the text: it is a signal, not proof, and when in doubt the policy asks for approval. ↩
- It applies to Claude Code. Codex has no equivalent hook, and its actions do not go through the policy. Approval is given only in the browser and is bound to the input hash. ↩
- In the runtime and the Cloud IDE, the receipt is signed with Ed25519 and chained to the previous one. In the demo, receipts are rehearsal receipts, chained by SHA-256 and unsigned. A receipt proves what went through the policy, not the absence of what did not. ↩
THE DECISION IN ONE PICTURE
Request, policy, one of three exits, receipt.
It is the path of every Claude Code action in the Cloud IDE and of every request in the runtime console. The demo shows the same path as a rehearsal.
Fail closed: if Harness cannot decide, the answer is no.
No answer from the policy, a timeout or an error along the way, and the action does not run. Nobody has to remember to block: blocking is the default, and allowing takes a written rule or a person. When the path itself goes down there may be no receipt left, but a missing receipt never turns into permission.
WHERE HARNESS FITS
It complements the agents you already use.
Tools like Claude Code and Codex write code and run agents for developers. Harness is the governance layer for business teams: it decides who may act, with which data and with whose approval, and leaves the receipt.
- Portable skills
- The same open skill format, installed from Agents & Skills.
- Rehearsal before real data
- Workflows tested on synthetic data from Evidence before touching the customer base.
- Model neutral
- The policy and the receipt do not depend on the model running the agent.
LIVE
What went live, with dates.
Each entry carries the commit in the Reveno repository and, where it exists, the public proof anyone can check.
Main protected in git itself
Cloud IDE 0.6.0. When an agent (Claude Code or Codex) tries to change a protected branch, the workspace git asks Harness and only proceeds with your recent approval. Approvals wait 9 minutes and alert you in the browser; receipts download as a bundle with an offline verifier.
Sessions on branches, each with its own worktree and agent
Cloud IDE 0.5.0. One session per branch, with its own worktree and the chosen agent; an agent on main can open and follow the others. A commit on main asks for approval, and merging a branch with the button becomes a signed receipt.
Trust page and first resolved advisory
Isolation layers, what the policy guarantees and what it does not, public receipt keys and advisory HX-2026-01: the hook could let an action proceed when it ran out of time; it now has a hard deadline and blocks on any failure.
Policy before every Claude Code tool
Cloud IDE 0.4.0. Every tool, MCP included, goes through the policy before it runs: it proceeds, waits for a person in the browser or is denied. Approval is bound to the input hash and works once.
A signed receipt for every agent action
Every tool Claude Code uses becomes a chained record signed with Ed25519, with secrets masked, and the IDE Receipts panel checks hash, chain and signature.
Catalog skills in the workspace and a usage bar
reveno-instalardownloads the public Agents & Skills package and only extracts the skills folder. The IDE gained 5 seats, a bar with memory, disk and seats, and extra keys on mobile.Runtime with a local 7B model
The console moved to the server's local AI, with no external key and no per token cost. Every answer still requires review.
Cloud IDE by invitation
One shell per person in the browser, under gVisor, to use their own Claude Code or Codex account. Single use invitation, password stored with scrypt and a session that survives a closed tab.
Runtime console and public keys
Reveno text skills with policy, approval with segregation of duties and an Ed25519 receipt checked in the browser.
Runtime v1 on the server, in rehearsal mode
Its own service with identity by workspace token, a worker with limited memory and time, and a kill switch. The second rehearsal passed, with the receipt chain checked outside the server.
Next
- Allowed internet destinations and a disk quota per workspace
- Credential custody and governed connectors
- Real cost per owner and per workspace
- External security review before opening to clients
Today, with Reveno
- Owners, policy, approval and records installed in the workflows you already use
- Diagnostic in 10 days, implementation in 6 weeks, monthly governance
HOW TO GET IN
From request to first receipt, in four steps.
Access opens gradually: there is a single server and the external security review has not happened yet.
- 01
Join the waitlist
Through the Reveno form, with the case you want to run. No card and no commitment.
- 02
A 15 minute call
Reveno walks you through the demo and agrees with you on what goes into the workspace and what stays out.
- 03
Single use invitation
The link is handed only to you and works once, for 7 days. There you create a password of 12 characters or more.
- 04
Workspace and first receipt
Sign in to the Cloud IDE, open Claude Code or Codex with your own account and make the first request. The Receipts panel shows the decision and checks the signature.
Seats on the server
5
workspaces running at the same time; the sixth sees "no seat"
CPU per workspace
1 vCPU
container ceiling
Memory
2 GB
container ceiling, up to 512 processes
Disk
8 GB
on screen warning from here; no hard quota yet
No terminal open
30 min
the workspace shuts down; files stay
Price
Not published
the AI model runs on your account; Reveno neither pays for nor stores anyone's key
QUESTIONS
Before you join the waitlist.
Does the demo execute anything?
No. It runs entirely in your browser, makes no network call and every receipt records that no action was performed. The policy decides as if the runtime existed.
What happens to the data in a request?
In the runtime, the request is kept only until the run ends and the answer for 7 days. The receipt stores only fingerprints (HMAC) of both, never the text. A request with personal or sensitive data waits for an approver before it runs. For client data at scale, Reveno installs the governance in the environments you already use.
Which AI model does it work with?
With the person's own. In the Cloud IDE, everyone signs in with their own Claude Code or Codex account, and Reveno neither pays for nor holds anyone's key. The console uses the server's local AI (a 7B model, with no external key and no per token cost), and every answer needs review. The policy and the receipt are model neutral, and skills follow the open SKILL.md format.
Is the receipt signed?
In the runtime, yes: every receipt is signed with Ed25519 and chained to the previous one. Public keys live at /.well-known/harness-keys.json and the console checks the signature in your browser. In the demo, receipts are rehearsal receipts chained by SHA-256.
How much does it cost?
Harness pricing is not published yet. Early access is by invitation, one workspace at a time, and the AI model cost sits on the account of whoever uses it. To install owner, policy, approval and records in the flows your company already uses, Reveno starts with the Diagnostic.
Where are the security limits written down?
On the trust page: isolation layers, what the policy guarantees and what it does not, the public receipt keys, resolved advisories and how to report a vulnerability.
EARLY ACCESS
Open by invitation, one workspace at a time.
The Cloud IDE opens one isolated shell per person, with policy deciding before every Claude Code action and a receipt for each one, and the runtime runs Reveno text skills with human approval and a signed receipt. Third party connectors and credentials come next. Nothing runs on these pages.