THE SCHEMA
BUILT FOR AGENTS.
CMP42 is a headless CRM in three deliberately small parts — Context, Models, Processes — exposed through MCP tools and a fast, boring REST + GraphQL API. Built knowing the primary reader will be an agent, and the primary writer a human helped by an agent.
CRM IS A GRAPH PROBLEM NOW.
Ask a sales manager how they check the pipeline. A year ago the answer was “I open HubSpot.” Today the honest answer is “I ask my agent to summarise it, and if the summary looks off, I open HubSpot.” The workflow has moved. The database has not.
HubSpot, Salesforce and Pipedrive were designed on the assumption that a human would open the UI. Every field, every pipeline, every dashboard exists because someone was going to click on it. When an agent reads the same schema, it treats it like a chat history: it hallucinates the missing bits, invents fields that were never declared, fabricates a stage that is not in the pipeline. Not because the model is bad — because the schema was not built for the reader.
CMP42 is what happens if you build a CRM for that reader. Structured context in, structured actions out, every write attributable. We ship no proprietary AI layer: CMP42 gives your agent grounded context, and your agent produces the insight. Read the full argument →
CONTEXT. MODELS. PROCESSES.
Three deliberately small parts that build on one another. All three are included in every plan.
The connected object graph
People, content and knowledge — contacts and deals, but also documents, meeting recordings and playbooks — as one graph. Every edge is typed, every field addressable by URI, every write versioned.
Agents read it like a database, not like a chat history.
The ontology you version like code
Lean defaults for contact, company, deal, task and activity. Extend them in TypeScript-shaped YAML, keep them in git, migrate with a single command, validate at write-time.
Every model becomes MCP tools automatically.
Agent-plus-human workflows, observable
Named sequences of steps — some run by agents, some by humans with approval hooks. Every transition logged, every run replayable against a past snapshot, token spend visible per run.
Audit an agent’s work like a git branch.
FOUR THINGS WE SHIP.
What we refuse to build is deliberate: no forced UI, no “AI insights” layer, no 200 fields on a contact. What we deliver instead is just as deliberate.
MCP tools, day one
Every model becomes a set of MCP tools automatically. Point Claude at your workspace and it can read, write and plan against your real context in one command.
REST + GraphQL parity
For non-MCP clients, every operation is exposed through REST and GraphQL. Same auth, same audit trail. Ship the agent of the month against the same surface.
Headless
No forced UI, no drag-and-drop pipeline builder, no dashboards that go stale on Tuesday. The admin surface is a data table and a migrations CLI.
Self-hostable
Managed by us in the EU, or run it yourself on your Postgres and object store. Same binary, same models. Your data, your call.
AGENTS ON TOP.
THE GRAPH UNDERNEATH.
CMP42 is layered so that every client — an MCP-native assistant, a Slack bot, the internal app your engineers ship next month — meets the same models, the same rules and the same audit trail. Read the table from top to bottom.
| Layer | What lives there | Read more |
|---|---|---|
| Clients | Claude Desktop, Claude Code, Cursor, Cline and custom agents built on the MCP SDK — plus Slack bots, internal apps and anything else that speaks HTTP. | Integrations |
| Surfaces | MCP tools, generated per model. REST and GraphQL with full operation parity. One auth model and one audit trail across all three. | MCP tools · API |
| Core | Context — the typed, versioned object graph. Models — your git-versioned ontology, validated at write-time. Processes — observable workflows of agents and humans. | Context · Models · Processes |
| Operations | A data table and a migrations CLI. No dashboards until you write them yourself. | Headless |
| Runtime | Managed by us in the EU, or self-hosted on your Postgres and object store. Same binary either way. | Self-hosting |
One request, end to end
- A partner asks their assistant to draft the renewal note for an account.
- The assistant calls MCP tools such as
get_contextandlist_relatedto walk the graph — contacts, the deal, the last documents. - It writes a draft back with
write_note. The write is validated against your models; if a value does not fit, the error tells the agent how to correct itself. - A process routes the draft to a human approval step before anything goes out.
- Every read, write, approval and token spent is recorded against the agent or human that produced it — and the run can be replayed later.
SHAPE THE PRODUCT BEFORE GA.
Twelve companies spend six months with us in Alpha, starting October 2026, for a one-time €6,900. Direct access to the founder, a monthly co-design call, priority feature requests — and the right to propose and name one MCP tool that ships to every future customer. Applications close December 2026.