AI that runs across every module, not next to one.
Generative answers grounded in your own graph. Agents that take action across Sales, projects, support, docs, and financials in one move. A model abstraction that never locks you to one vendor.
Foundation — A live product capture for this area is not published yet — we only show screenshots of the running app.
What AI does for you
Generative answers, grounded
Every entity exposes an `ai_context` serializer. Agents read the actual record, not a vector summary, when grounding matters.
Agentic actions
Agents create records, draft documents, send messages, and move work. Permissions and the JWT roleId scope what each agent can do.
Model abstraction
Anthropic Claude is the primary; OpenAI is the secondary. Add or swap providers in one place — callers never know which model answered.
Audited and reversible
Every agent action is logged and reversible. Customers see exactly what the AI did and why. No silent automation.
AI agents need context. Cyril gives them all of it.
An agent that can only see a CRM has thin context. An agent that can read Sales + projects + support + docs + financials in one query has the context to actually decide. Cyril’s graph is what makes the AI useful.
- Single auth model across modules
- `ai_context` on every entity — never just an embedding
- MCP server (`packages/mcp`) exposes tools to external agents too
Three modes, describing how hands-on you choose to be.
Platform is the full application, driven by you — nothing happens that you did not do. AI is where you ask, and it plans, shows the plan, asks approval, and acts where you can watch. Agentic, where you hand a goal over entirely, is built but described as coming soon rather than generally available. The modes are about your involvement, not about how much AI is present: the layer is there in all three.
- Platform: you drive, nothing happens on its own
- AI: it plans, shows the plan, and acts with approval
- Agentic: coming soon — do not plan around it today
Common AI questions
The model abstraction layer is designed for it. Adding a new provider is a single adapter file. At launch we ship Anthropic Claude as primary and OpenAI as secondary; self-hosted (Llama, Mistral) is on the roadmap.
Permissions and the JWT `roleId` scope every agent action. Risky actions require human approval. Every action is logged, auditable, and reversible. Customers see what the AI did and why.
No. The AI is a caller subject to the same permission model as a person, not an exception to it. That is the property that makes it safe to point it at everything — and it is also why a broad question returns less for a narrowly-permissioned user.
Every agent action writes to the same audit log as a human one, with the actor and the authority it acted under, and goes through the same transactional, soft-deleting write path — so the inverse already exists. There is no second, weaker write route for automated callers.
Be one of the first to use Cyril.
Join the waitlist for early access. We'll only email you when there's something real to share.