For people building real software with Codex and Claude Code
Your AI-built product worked. Then the endless fixes started.
Consensus Logic is a Windows app that works out what is actually wrong. Then Codex and Claude each plan the work, check each other's plan, and only the steps you approve get run.
No API keys, no extra AI bill. It runs on the Codex and Claude subscriptions you already pay for.
Nothing to download yet. The first version does documentation and specification tasks; code tasks follow.
The fix loop
The four-hour fix. Then another four-hour fix. And it is still broken.
The model says it has fixed the problem. Tests may pass. The visible problem remains. Another repair begins, then another. By the time the real issue is found, a day, several days, or an entire week may have disappeared into the loop.
The cost is not just time. It is lost confidence in the code, uncertainty about what changed, wasted model quota, and the feeling that the project is no longer under control.
You are not shopping for another coding assistant. You are trying to recover control of something you have already put days or weeks into.
Illustrative record. The pattern is real; we have lived it.
Already in the loop
Stop fixing the symptom. Recover the system.
When a build has already gone wrong, another giant repair prompt is not the answer. Codex and Claude read the code separately, argue about what is actually broken, work out what the whole thing is supposed to do, then repair it in the order the system depends on, checking each step before the next one starts.
Narrated, with captions. Under a minute. It plays only when you press play; nothing on this site moves on its own.
Narrated, under a minute. Captions on screen.
What you get
At the end of every task, a plain answer to five questions.
Every task closes the same way: what did I get, exactly where is it, how do I try it, what is still wrong or unknown, and how long did it take. The Plain English tab and the Technical tab show the same evidence. Never a different story.
Illustrative record. Real screens replace this when the first working version is tested.
- What did I get
- Section 3 rewritten from the current source, with its terms matched to sections 1 and 2. The check passed: no section contradicts another.
- Exactly where is it
docs/customer-guide/03-getting-started.md, revision 4. The previous revision is kept beside it.- How do I try it
- Open the file. Every cross-reference resolves. The glossary terms match section 1.
- What is still wrong or unknown
- Screenshots were not regenerated. Attachment input is Not run until the full workflow version.
- How long did it take
- 1 h 42 min elapsed. 38 min active, 51 min waiting on the providers. Overlap disclosed.
The way back
How it gets you from a broken build to a checked fix.
It reads your project, gets two plans, makes them challenge each other, brings the real disagreements to you, then runs small checked steps and records what actually passed.
- 01
It reads your project
What is actually in the project now, what changed, and what the evidence says works. Before anyone plans anything.
- 02
Two independent plans
Codex and Claude get the same evidence and each write a plan without seeing the other's. Neither sees the other's work until both are done.
- 03
Challenge
Each plan is cross-examined by the other for omissions, dependencies, size and proof. A clean pass is allowed. Manufactured objections are not.
- 04
Your decision
A real disagreement comes to you with the trade-off stated. Nothing is weakened to make the two agree. You resolve it.
- 05
Small steps, each with a limit
Only after you have seen the plan in plain English and separately authorised it does work run, one small step at a time, each with a size limit and a check, inside the budget you set. The app calls that budget your allowance.
- 06
A record you can check
Every step ends with what changed, what passed, what failed and what is still unknown. A model's “done” is not evidence. The controller reads the actual output.
Two plans into one programme
Written separately
Same evidence, no sight of the other plan.
Written separately
Same evidence, no sight of the other plan.
Challenge
Cross-examinationOmissions, dependencies, size, proof. Clean pass allowed.
Real disagreements
Trade-off stated. You resolve it. Nothing is weakened to agree.
Agreed
The programmeSteps, dependencies, limits, checks. Plain-English summary first.
It remembers your project
What exists, what was examined, what was decided and what is still open is kept as a durable record, so a new task starts from what is known, not from a blank chat.
Every plan is checked against the real code
Before a plan is trusted, the AI that did not write it reads the source the plan depends on and reports what it actually found. A plan built on a misreading is caught before anything runs.
Nothing runs until you say so, and every step leaves a record
Only after you have seen the plan in plain English and authorised it does work run, one small step at a time inside the budget you set. Each step closes with what changed, what passed, what failed and what is still unknown. A model saying “done” changes nothing; the app reads the actual output.
Will the next model make this unnecessary?
Better models do not remove the problem.
Consensus Logic is pro-AI. Claude, Codex and the coding models that follow them are extraordinarily capable, and that capability is the reason this product exists.
But capability is not infallibility. A new model release does not remove the need for independent review, limits the AI cannot cross, and evidence. A model can still:
- 01
Misunderstand a system it has only partly read.
- 02
Optimise around the wrong assumption, confidently.
- 03
Repair the symptom rather than the cause, repeatedly.
- 04
Report success before the real application works.
AI is powerful enough to build serious products. That is exactly why serious work needs a layer of control around it.
Proof
Show, not claim. Including what is not done.
Nothing is proven yet. There is a list of 58 things the app must do, and each one is marked as it is tested in the real, installed app. The Proof page shows the whole list with its real status, including what has failed or has not run. Right now almost all of it has not run, and the site says so.
From the list of 58
| Id | Scenario | Status |
|---|---|---|
| A016 | Both planners write their plans without seeing each other's, from the same evidence | Not run |
| A057 | A real disagreement becomes your decision, with the trade-off stated | Not run |
| A058 | Execution cannot be authorised until the plain-English summary has been shown | Not run |
| A011 | Every AI call reserved against the budget before it starts, settled after | Not run |
| A036 | A model's “done” and a green test count cannot unlock the next step | Not run |
| A017 | Stop starts nothing new immediately; if it cannot confirm the work stopped, nothing else runs | Not run |
| A031 | Microphone input | Not implemented |
What it will tell you it cannot do yet
| Capability | Arrives in | Status |
|---|---|---|
| Documentation tasks end to end through the installed app | First working version | Not run |
| Attachments, microphone, both report tabs, automatic supervision, multi-project | Full workflow version | Not run |
| Code and mixed tasks, with release proof | Code tasks version | Not run |
| Team administration, SSO, cloud hosting | — | Not implemented |
| macOS and Linux | — | Not implemented |
Updated as each version is tested. The site is allowed to say only what a passed test proves. Teams and other platforms each have a waitlist.
The commercial decision
What did the last failed repair loop cost you?
Not in tokens. In hours, days and confidence. If Consensus Logic prevents one lost day, rescues one broken development cycle, or stops one week of circular repair work, the annual licence should already feel inexpensive.
Planning is free. You pay only when you press Start development to run the plan.
Does it use the OpenAI or Anthropic API?
No. It drives the command-line tools that come with your existing Codex and Claude subscriptions. There is no separate Consensus Logic inference bill and no API key to fund.
Do I need both subscriptions?
To use both AIs, yes: supported access to both Codex and Claude. The exact supported plans and command-line prerequisites are listed on the Download page and kept current.
Does it give me extra Codex or Claude quota?
No. Vendor usage limits still apply. Consensus Logic uses the access you already have. Its purpose is to keep less of that access going into blind repair loops; actual usage depends on your project.
What do I need to run it?
A Windows 11 machine, and the Codex and Claude command-line tools signed in. Windows is the launch platform. macOS and Linux are not supported yet; there is a platform waitlist.
Where does my project data live?
On your machine, in your local application data folder. Projects, plans, records and reports stay there. The AI providers receive project context under their own terms, because they are the models doing the work. The privacy page says exactly what leaves the machine.
The waitlist
Get told when there is something to download.
Nothing is downloadable yet. The build log is the first thing you will get, and it applies the product's own standards to the product itself.
There are no dates on this site, on purpose. A version ships when it passes its tests in the real app, not when a calendar says so.
Draft: waitlist endpoint not wired yet