One control point for enterprise data and AI, running where your data lives.
Data.co Control connects your systems, labels what is sensitive, decides what every user, agent, and model may do with it, and serves private models behind one gateway. The runtime is installed in your environment. Our hosted plane only handles your account and license.
Two planes, one hard boundary.
Data.co Control is split in two. The Control runtime (connectors, catalog, classification, policy engine, gateways, model serving, and the audit log) is installed inside your infrastructure: on-premises, in a private cloud account, across a hybrid footprint, or fully air-gapped. Every prompt, document, embedding, and database record is processed there.
The hosted account plane at app.data.co handles accounts, licensing, billing, and, only if you opt in, fleet administration such as version visibility across deployments. It never receives prompts, responses, documents, embeddings, or records, and the runtime does not need it to make access decisions. If the link to app.data.co goes down, or never existed, your governed AI keeps running.
Five modules, each with a defined job.
Connect, Govern, and AI run in the customer-local runtime. Deploy and Manage describe where that runtime lives and who operates it.
| Module | What it does | Where it runs |
|---|---|---|
| Connect | Connector runtime, data catalog, classification at ingest, lineage, scheduled sync and change data capture | Your network |
| Govern | Identities, deterministic policy engine, enforcement points, approvals, DLP, append-only audit log | Your network |
| AI | OpenAI-compatible AI gateway, private model serving, permission-aware RAG, MCP gateway for agent tools | Your network |
| Deploy | Reference architectures and hardware for on-prem, private cloud, hybrid, and air-gapped installs | Your facilities or cloud account |
| Manage | Optional managed operation: patching, monitoring, backup, and recovery of your deployment | Your environment, operated by our team under your access rules |
The governance capabilities inside Control.
Each capability has its own page. They share one identity model, one policy engine, and one audit log.
AI Governance
Turn a written AI governance program into versioned policies that run on every request and produce their own evidence.
02AI Identity & Access Control
Distinct identities for users, service accounts, agents, and models, so an agent never sees more than the user it serves.
03Policy Engine
A deterministic decision point that returns allow, deny, or require approval, with the policy ID that decided it.
04AI Agent Security
Limit what a manipulated agent can reach: scoped data, governed tools, and credentials the model never holds.
05MCP Security Gateway
Filter which MCP tools each identity sees and validate every tools/call argument against policy before it runs.
06Data Loss Prevention
Classification-aware egress rules for prompts, tool calls, exports, and write-backs, with external models off by default.
07Audit Logs & Compliance
An append-only record of every request, decision, source, and tool call, kept locally and exportable as evidence.
08Human Approval Workflows
Hold destructive or high-value actions for a named approver, with a timeout that defaults to deny.
09AI Monitoring
Content-free usage, latency, model health, and policy-outcome telemetry that stays inside your environment.
What happens when someone asks a question.
A single request through the AI gateway touches every module in the runtime, in a fixed order.
- 1
Authenticate
The gateway resolves the user, and the agent or application acting for them, to identities Control knows about.
- 2
Decide
The policy engine evaluates the request against classification and context. No model participates in this step, and it fails closed.
- 3
Retrieve + infer
Retrieval returns only content the requesting identity can open in the source system; the prompt goes to an approved model endpoint.
- 4
Record
Identity, decision, sources, model, and outcome are written to the append-only audit log in your environment.
What app.data.co never sees.
- Prompts and model responses
- Documents and file contents
- Embeddings and vector indexes
- Database records
- Connector and model credentials
- Audit log contents
- Policy decisions for individual requests
- Classification results
How teams usually try to govern AI, and where each approach stops.
None of these approaches is wrong for every situation. The question is whether the control holds when a model misbehaves, a credential leaks, or an auditor asks for proof.
| Approach | What it does well | Where it stops |
|---|---|---|
| Instructions in system prompts | Fast to change; shapes tone and scope of answers | Probabilistic. Retrieved content or a user can override it, and it leaves no decision record |
| In-house proxy and scripts | Fits one application exactly | Each team rebuilds identity, classification, approvals, and logging; rules drift between apps |
| A cloud provider's AI tooling | Integrated with that provider's models and storage | Assumes data and inference in that provider's cloud; harder for on-prem, multi-cloud, or air-gapped estates |
| Data.co Control | One identity model, deterministic policy, and one audit log across models, agents, tools, and sources | Governs traffic routed through its gateways and connectors; other paths still need network and endpoint controls |
Three common starting points.
Control is modular, so the first deployment is usually shaped by the most pressing risk rather than by a full platform rollout.
- 1
Agents already in production
Put the MCP gateway and AI gateway in front of existing agents, register their identities, and start with policies that mirror what they do today. Tighten scope once decisions are visible in the log.
- 2
A regulated knowledge assistant
Connect the document sources, classify at ingest, and serve permission-aware retrieval through a private model. The first policies usually keep PHI, PII, or privileged content on private endpoints.
- 3
A governance program needing evidence
Inventory sources and model endpoints, translate written rules into tested policies, and schedule audit exports. Enforcement can be turned on per gateway as owners sign off.
Built to run with no outbound connection.
Licenses validate offline, policy decisions never call out, and external model endpoints are off by default. Air-gapped is a supported deployment target, not an afterthought.
Questions buyers ask
What exactly is hosted by Data.co?
Only the account plane at app.data.co: user accounts for your administrators, licensing, billing, and optional fleet administration. It does not receive prompts, documents, embeddings, records, or audit log contents.
Do we have to adopt every module?
No. Teams often start with Connect and Govern in front of an existing model, or with the AI gateway and policy engine in front of agents they already run. The modules share identities and the audit log, so adding one later does not mean re-integrating.
Can we use external model providers?
External model endpoints are off by default. You can enable specific ones, and policies then decide which data classifications, if any, may be sent to them. Private models served with vLLM or Ollama on your hardware are the default path.
Who operates the runtime?
Your team, or ours under a managed services agreement. In both cases it runs in your environment under your access controls, and managed operation does not change what data leaves the boundary.
How mature is the product?
Data.co Control is early. We scope each deployment against your systems and confirm which connectors and capabilities are available for it before any commitment, rather than publishing a blanket compatibility list.
How is the license enforced without calling home?
Licenses are signed files verified offline by the runtime against a public key. A license names the licensed modules and expiry, with a grace period, so renewal is a file replacement rather than a network dependency.
See it against your own use case.
A 30-minute technical walkthrough, policy simulation on your scenarios, and a deployment recommendation for on-prem, private cloud, or hybrid.