Propose
Each agent works under a scoped, revocable credential. It can propose changes. It can't approve them.
Not even its own. Credentials are short-lived and scoped, and revocation takes effect immediately.
By Neural Forge
Agents propose.
Verifiers decide.
Tesseract puts a deterministic gate in front of every AI-written change. Real tools check each change for code health, contract stability and compliance, and it ships only when all three pass. On Enterprise Hybrid and Platform, every decision is signed and kept on a record you can verify yourself.
- Binary gate
- No override switch
- Signed results
- Your code stays with you
The problem
Your agents write code faster than anyone can review it.
AI coding agents turn a ticket into a pull request in minutes. Review turns into skimming. Adding an AI reviewer means one model grading another, and both can be confidently wrong. A prompt-injected agent can report that everything passed.
None of this fails loudly. Type safety slips a little, an API field quietly disappears, a debug log ends up in production. Six months later an audit asks who approved a change, and nobody can say.
Why now
Agents now open pull requests on their own
Coding agents have moved from suggesting a line to writing whole changes.
Review doesn't scale with them
More code arrives every week. The number of people who can review it doesn't change.
Auditors are starting to ask
Who approved an AI-written change, and on what evidence?
The answer
Take the judgment call out of the merge path.
Tesseract never asks a model whether a change is good. It runs the same tools your best engineer would run, every time, on every change, and applies one rule that isn’t negotiable: all three checks pass, or the change doesn’t commit. No weighting, no partial credit, no override flag.
Check 1 of 3
Code health
Type errors must be zero. Lint findings and complexity are held to a baseline calibrated on your own codebase.
Check 2 of 3
Contract stability
The interfaces other code depends on are compared with the last approved version. Unexplained breaking changes fail the gate.
Check 3 of 3
Compliance risk
Policy rules catch committed secrets, unsafe type escapes and leftover debug output, plus any rule packs you install.
Pass all three, and the change ships. Miss one, and it goes back with the exact reason.
How it works
Five stages from proposal to trust.
Here is what happens between an agent writing a change and that change reaching your main branch. Every change must pass all three checks before it ships.
Principles
Built so the safe path is the only path.
Four rules the system is built around. They're part of how it works, not settings someone can switch off.
Trust output, not claims
Nothing an agent says about its own work counts. Only a verifier's result does.
Same bar on every plan
Lighter plans get lighter checks. The pass bar itself never drops.
Monitors can't act
Kyrios has no block or revert function. It isn't switched off. The capability simply doesn't exist.
Automation is earned
Every service starts at Tier 0. Each step up needs a named person's sign-off, kept in an audit log that can't be edited.
Products
One gate, three ways to run it.
The same pass bar on every plan. What changes is where the checks run and how much of the system you take on.
Tesseract
In developmentLite
For small teams and solo builders
Checks run in your editor while your AI assistant writes, and a real gate runs before you commit.
- Early hints as code streams in. Advice only, never the verdict
- The authoritative check on save and before commit
- Bring your own model or API key
- Your code never leaves your machine
Tesseract
Design partnersEnterprise Hybrid
For regulated and security-conscious teams
The gate runs on developer machines and in your CI, signs every result with your key, and blocks merges that can't prove they passed.
- A local daemon at commit time and in CI
- Results signed with a key that stays on your side
- A required CI check that blocks unproven merges
- Cloud, on-prem or hybrid. Starts in shadow mode
Tesseract
Design partnersPlatform
For organisations running agent fleets
The full system, deployed just for you: scoped agents, a complete change history and drift monitoring across every service.
- A scoped agent fleet, included
- Every change traceable back through its gate results
- Drift monitoring across all your services
- The autonomy ladder, with named sign-offs
Compare the options
| Feature | Lite | Enterprise Hybrid | Platform |
|---|---|---|---|
| Binary three-check gate | Yes | Yes | Yes |
| Where checks run | Your editor | Developer machines and CI | Your dedicated deployment |
| Signed attestations | No | Yes | Yes |
| CI merge block | No | Yes | Yes |
| Vector Packs | Lite-sized packs | Yes | Yes |
| Kyrios drift monitoring | No | No | Across services |
| Agents | Your own assistant | Your own agents | Scoped fleet included |
| Autonomy ladder | No | No | Yes |
| Pricing | At launch | On request | On request |
Deployment options
Signed proof behind every merge, on your terms.
Run it in your cloud, on-prem, or across both. Nothing blocks until you've seen what it would have done.
Start in shadow, go live when you're ready
By default Tesseract measures and records but never blocks. Once you've seen what it would have done on real work, switch it to live. Without a valid licence it drops back to shadow mode and says so loudly. It never fails silently.
You choose what an outage means
If the control plane can't be reached, the CI gate reports a degraded result and warns loudly. Turn on strict mode and any outage blocks the merge, for pipelines where a missing check must never mean a pass.
Your keys, your code, your record
The signing key is created and kept on your side, and we only ever see the public half. Source code stays in your environment: the control plane receives signed results, not code.
Cloud
Daemon on developer machines and CI runners. Control plane in your cloud account.
On-prem
Everything inside your network, including the control plane.
Hybrid
Daemon and CI on-prem, with the control plane wherever your security team prefers.
Vector Packs
Rules for your domain, held to the same bar.
Vector Packs add domain rules to the gate. Each pack is readable, version-pinned and reviewable like code. It can make the gate stricter. It can never loosen it.
Available
FinTech Risk Invariants
Rules for code that handles money, accounts and financial risk.
Available
SOC 2 Secure Coding Baseline
Secure-coding checks that line up with SOC 2 control expectations: committed secrets, type safety, debug output.
Available
Data Integrity Invariants
Rules that protect customer records and the integrity of the data your services keep.
On the roadmap:
- Healthcare data (PHI)
- Payment cards (PCI)
- Privacy (GDPR)
- Accessibility (WCAG)
Design partners decide the order.
The SOC 2 pack supports your controls. It doesn’t certify you or replace an audit.
Security and trust
Designed so you don’t have to trust us.
Every claim Tesseract makes about a change comes with something you can check yourself: a signature, a tamper-evident record, a pinned tool version.
One client per deployment
Each client gets its own server, agents and history. There's no shared system where your data sits next to someone else's.
Least privilege for agents
Agents authenticate with short-lived, scoped credentials. Revocation takes effect immediately, not when a credential expires.
Signed, tamper-evident, exportable
Ed25519 signatures, with the private key kept by you. Any gap or rewrite in the record is detected, and exports verify offline with your public key.
Logs that can't be edited
Tier changes and drift reports go into audit logs with no edit or delete function. Changing history would take new code, not a permission someone forgot to remove.
Safe defaults
Encrypted connections only, credentials never travel in URLs, and a pull request can't swap the key that's used to check it.
How we handle personal information is set out in our privacy policy and PAIA manual. Your security team can go through the full trust model with our engineers on a call.
Languages
Supported languages
A language counts as supported when its checks run on pinned tool versions, give the same answer every time, fail safely if a tool breaks, and have been calibrated on real code. Here's where each one stands.
| Language | Status | Frameworks |
|---|---|---|
| TypeScript / Node.js | Available | Node.js services |
| JavaScript / TypeScript web | Planned | React, Next.js, Vue, Nuxt, Angular, Express, NestJS |
| HTML / CSS | Planned | As part of web projects: accessibility, markup and stylesheet checks |
| Python | Planned | Django, FastAPI, Flask, SQLAlchemy |
| Java | Planned | Spring Boot, Jakarta EE |
| C# | Planned | ASP.NET Core, Entity Framework Core |
| Go | Planned | net/http, Gin, gRPC |
| Ruby | Planned | Rails, Sinatra |
| C++ | Planned | Public headers, Qt, gRPC |
Contract-stability checks for services that publish OpenAPI, protobuf or GraphQL contracts are planned alongside the languages above. If your stack isn’t available yet, tell us. Design partners set the order.
Design partner programme
Help shape the gate you’ll rely on.
We're working with a small group of teams that ship AI-written code to production. You get early access and direct support from the engineers who build Tesseract. We learn from running it on real codebases, inside your environment, and from your honest feedback.
You get
- A shadow-mode run against your own commit history, with a report of what the gate would have caught.
- Calibration tuned to your codebase.
- Direct access to the engineers building Tesseract. No account managers.
- A real say in which languages and rule packs come next.
- Design-partner pricing.
We ask
- One service or repository to start, in a supported language or one you want us to prioritise.
- A named technical contact and roughly an hour every two weeks.
- Honest feedback, including when something doesn't work.
A good fit
Teams that already use AI coding agents in production work, that answer to auditors or security reviews, and that want evidence rather than reassurance.
How it runs
Apply
Tell us about your team and how you use AI coding tools.
Intro call
A conversation with the engineers who build Tesseract.
Security review
We walk your security and platform leads through the trust model.
Shadow run
Tesseract runs on your own history, in your environment. Nothing blocks.
Report
What the gate would have caught, and a decision on going live.
FAQ
Questions, answered.
No. Agents propose changes, but the pass or fail decision comes from deterministic tools: compilers, linters, contract comparison and policy engines. The same commit always gets the same result.
We’re choosing our first design partners.
We’re onboarding a small number of teams who ship AI-written code to production and want proof behind every merge. You’ll work directly with the engineers building Tesseract, and help decide which languages and rule packs come next.