SELL THE AGENT
THAT KNOWS THE ACCOUNT.
You are already selling AI to your clients. Sell them the version that grounds on their actual accounts — not the demo that hallucinates a competitor.
DEMOS ARE EASY. GROUNDING IS THE WORK.
The first AI project for a client usually looks great in the pitch. A capable model, a tidy prompt, a few sample records. Then it meets the client’s real data: a CRM configured over years for people clicking through screens, fields nobody remembers adding, knowledge spread across drives and wikis.
Now the agent answers confidently and wrongly. It invents a field. It names a competitor as a customer. And the reputation on the line is not the model vendor’s — it is your agency’s.
The durable part of an AI engagement is not the prompt. It is the structured context underneath it: an ontology that describes the client’s business, a graph an agent can read, and a trail that shows what the agent did. That is what CMP42 is for.
ONE ONTOLOGY PER CLIENT.
A per-client ontology
Every client’s business is different: distributors here, franchise locations there. Describe it in Models — TypeScript-shaped YAML, versioned in git, migrated with one command.
Migrations are backwards-compatible, so a schema change for one client does not surprise the agents already in production.
The model your client approved
CMP42 ships no proprietary AI layer. Your client decides which model they trust; CMP42 provides the context. Anything MCP-compliant connects — Claude Desktop, Claude Code, Cursor, Cline or a custom agent on the MCP SDK.
For everything else, REST and GraphQL expose the same operations. Bring your own model →
Delivery you can show
Wrap agent work in Processes with human approval steps. Every run records what the agent read, what it wrote and what it cost in tokens — per process, per run.
That turns “the AI is working” into something you can review with the client, line by line.
MANAGED OR SELF-HOSTED.
| Managed EU tier | Self-hosted | |
|---|---|---|
| Who operates it | protagx, EU-hosted | You or your client, on their Postgres and object store |
| Training on data | CMP42 does not train on your data | Nothing leaves the environment at all |
| Tracking | No third-party analytics, fingerprinting or session replay | Under the operator’s control |
| Availability | Alpha from October 2026 | Image published with Beta in Q1 2027; Design Partners get Alpha access |
Clients differ in how much control they need over their data. CMP42 offers the same product two ways: operated by us in the EU, or run on the client’s own infrastructure. Same binary, same models, identical MCP tool surface — so the work you do on the ontology carries over either way.
If you plan to run CMP42 for several clients, raise it on the 15-minute call — we will tell you honestly what the Alpha supports today.
Multi-workspace operation — one agency running several isolated CMP42 workspaces, one per client — is a standard feature from the Team tier upwards, with an Agency-Owner role that has cross-workspace access while each client sees only their own. White-label (running CMP42 under an agency brand with a custom admin-UI domain for clients) is not part of the Alpha; it is planned for GA at the earliest, and interested agencies can register the intent on the 15-minute call so we shape the rollout around real cases. A formal reseller programme with fixed terms is deliberately not defined yet — for the Alpha the Design Partner Program is the entry point, and reseller conditions after Alpha are negotiated individually against the shape of your book of clients.
A CLIENT ASKS FOR A SALES ASSISTANT.
From mapping session to reviewed output
A fictional agency and a fictional client. The steps show how the concepts fit together; they do not describe a delivered project, and interfaces will change during Alpha.
- The agency runs a model-mapping session with the client: which objects matter beyond contact, company and deal — say, distributors and product lines — and which fields the sales team actually uses.
- The agency declares those as Models, reviews them with the client like code, and migrates with one command. MCP tools for the new models are generated automatically.
- The client’s sales team connects the agent they already use; the agency builds a small custom agent on the MCP SDK for the team’s Slack channel. Both read the same graph.
- Writes that matter — notes on key accounts — go through a Process with the client’s sales lead as the approval step.
- In the monthly review, the agency walks the client through run history and token cost per process: what the agents did, what was approved, what it cost.
FROM PROMPT WORK TO CONTEXT WORK.
An agency that owns the context layer sells something more durable than a prompt: a documented ontology, a graph that any future agent can read, and processes a client can audit. When the client switches models next year, that work remains.
One caveat, in the spirit of an honest Alpha: the Design Partner Program asks for real usage on real data, not a demo tenant. If you cannot yet bring a client workload, Beta opens in Q1 2027 with a waitlist.
Shape the tool surface your clients will use
Design Partners propose and name one MCP tool that ships to every future customer — and get a monthly co-design call on their roadmap.
→ Apply to be a Design Partner- Grounded answers with citations to the client’s own objects.
- An ontology the client owns, versioned and reviewable.
- Model choice stays with the client — no lock-in to an AI vendor through us.
- An audit trail per run, including token cost.
- EU-managed or self-hosted, depending on the client’s requirements.