Example scenario – illustration, not live data.
01CONSYRIUM AI
Claude, OpenAI, Gemini, Grok, DeepSeek, Kimi and other AI models think independently, challenge each other and reach a documented verdict with evidence. You approve – and take the decision where you work: as a report, in your tools or, with a connected GitHub repository, as a reviewed pull request.
02BEFORE THE COUNCIL
Create a project and describe the decision. Choose the models and up to twelve perspectives. You connect your own provider keys and set a hard token budget – CONSYRIUM AI sells no AI credits.
EXAMPLEQuestion: “Which database architecture should our SaaS product use?” – three models, twelve perspectives, a budget on your own keys.
03PHASE 1
Every model receives the identical brief and works on its own. Answers stay sealed until all are in – no model can see or copy another.
EXAMPLEModel A drafts a managed multi-tenant database, B a self-hosted cluster, C a hybrid – none of them sees the others.
04PHASE 2
All proposals are revealed at once and laid out side by side as equals. You compare the alternatives while the models ask each other precise questions.
EXAMPLESide by side: A is fastest to launch, B gives the most control, C balances both. B asks A how tenants are kept apart.
05PHASE 3
Roles such as architecture, security, cost and critic unfold around the models and test every proposal. Twelve perspectives are roles – not twelve extra models or votes.
EXAMPLEThe security role checks tenant isolation, the cost role compares running costs, the critic looks for weak spots.
06PHASE 4
A substantiated objection is highlighted and checked against evidence by independent models. Dissent is recorded – never smoothed away.
EXAMPLEOne objection stands out: in A, a misconfiguration could let tenants see each other’s data. Evidence: test results and provider docs.
07PHASE 5
Authors revise their proposals against the objections; compatible ideas can merge into a hybrid option. Differences that remain stay documented.
EXAMPLEA adds per-tenant keys and merges with C into “Hybrid A + C”. B’s cost concern stays documented.
08PHASE 6
Recommendation, independent scores, evidence and the minority position come together in one decision report. Nothing is built at this point.
EXAMPLERecommendation: Hybrid A + C – with scores for security, cost and time to launch, and the security objection on record.
09HUMAN APPROVAL
Approve, choose an alternative, ask for more debate, request changes or reject. A proposal with an open critical security objection cannot be approved. Without your approval nothing is built.
EXAMPLEYou approve Hybrid A + C – or ask for one more round on cost first.
10PHASE 7 · OPTIONAL, WITH GITHUB
Only if you want it and a GitHub repository is connected: after your approval the decision becomes tasks, code, tests and independent reviews – delivered as a draft pull request. Most decisions need no build; merge and deployment stay your separate decisions.
EXAMPLETasks: schema, per-tenant keys, migration tests – opened as a draft pull request in your repository.
11FROM QUESTION TO DECISION
Your keys, your providers, your approval. Every step leaves evidence you can follow.
Seven canonical stages. Modes – Independent Expert, Adversarial Review, Consensus – change how the council is orchestrated, not just the wording.
STAGE 1
Every model receives the identical brief and answers blind. No model sees another proposal before the reveal.
STAGE 2
All proposals are revealed side by side. Models ask each other precise questions and answer them.
STAGE 3
Up to twelve perspectives – architecture, security, cost, critic and more – review every proposal.
STAGE 4
Dissent is first-class. Substantiated objections are checked by independent models and may stay unresolved.
STAGE 5
Authors revise their proposals against the objections. Agreements and open disagreements are documented.
STAGE 6
Independent scores, hard gates and a separate majority vote produce a reproducible verdict – then you decide.
STAGE 7
Optional, with a connected GitHub repository: after your approval a task plan, sandboxed code, tests, reviews and a draft pull request.
HUMAN GATE
Approve, select an alternative, request more debate, request changes or reject. Proposals with a critical security objection cannot be approved.
Several AI models develop solutions, challenge each other and score them. You see the reasoning, open objections and results before you decide.
Up to six AI models work on the same decision.
Models challenge each other's proposals.
Solutions are measured against your confirmed hard requirements.
Scores, reasoning and unresolved objections side by side.
Final approval always stays with you.
CONSYRIUM AI is bring-your-own-key only. You connect your own API keys and pay the providers directly – CONSYRIUM AI sells no AI credits and has no fallback key.
Anthropic
Claude models
Your own API key
OpenAI
GPT models
Your own API key
Gemini models
Your own API key
DeepSeek
DeepSeek models
Your own API key
Z.ai
GLM models
Your own API key
Moonshot
Kimi models
Your own API key
Alibaba Cloud
Qwen models
Your own API key
xAI
Grok models
Your own API key
Most teams need the reasoned decision, not code. Take the verdict as a report or into your tools – a pull request is one option, only with a connected repository.
AVAILABLE
Recommendation, scores, dissent and evidence in the app
AVAILABLE
Export for your notes, wiki or own tools
AVAILABLE
Print-ready report to share or file
AVAILABLE
Summary with the report attached
AVAILABLE
Short verdict and link to your chat
AVAILABLE
A page in your team space
PLANNED
The report as a note in your vault
WITH A CONNECTED REPOSITORY
Optional build by the Execution Council – reviewed, tested, as a draft
You don't have to open the web app for every question. With Pro you start decisions from your development tool or your private Telegram chat.
MCP SERVER
The built-in MCP server connects CONSYRIUM AI to Claude Code, Claude Desktop, Cursor and VS Code.
TELEGRAM
Your own bot takes questions and sends the verdict back.
With a connected GitHub repository the Execution Council can build an approved decision – reviewed, tested and always as a draft for you to merge.
01
A human approves the verdict. Nothing is built before that.
02
The Execution Council turns the verdict into a validated task graph.
03
Code only runs isolated – in a sandbox without network access or in your repository's GitHub Actions.
04
Independent models review the change; findings stay visible.
05
A draft pull request opens in your repository through the GitHub App.
CONSYRIUM AI is free, open-source software. Download the source code from GitHub and run it yourself with Docker Compose.
Host it yourself or start online right away. AI usage always runs on your own API keys – CONSYRIUM AI does not sell AI credits.
€0
Free software · you cover server and AI costs
€0
on app.consyrium.com · AI costs billed by your provider
€4.99
per month incl. VAT · AI costs billed by your provider
Cancel Pro any time in the Stripe customer portal. In every option you pay AI API usage directly to your provider.
Describe an idea, an architecture question or connect a repository.