The Colrows platform

The semantic foundation for governed enterprise agents.

A semantic layer built from your data, governed access for every agent, and agents that run on your compute.

19
agent tools over REST and MCP
2
governed MCP endpoints
13
versioned semantic object types
0
raw SQL in an agent run’s grant
Architecture

Every layer, one governed loop.

Built from your data estate, shared by every agent, governed end to end.

Who asksPeople in Colrows appsAI Analyst & CuratorYour own agentsAny MCP clientYour applications
Colrows platform
Agent control plane
RegistrySpecs, versions
Review & approveApproval, evals
TriggersOn demand, scheduled
Answer checkAgainst evidence
Live agentsAI Analyst, Curator, yours
Grounded access
REST & MCPSame tools on both
Structured dataLive governed SQL
Unstructured dataCited, per-user access
Semantic layer
MetricsEntitiesBusiness termsRelationshipsVerified queriesAccess policies
Semantic inference engine
Extract signalsFrom every source
Reconcile claimsConfidence, conflicts
ProposeWith evidence
CuratePeople, Curator
Persistent memory
Agent memoryPrivate per user or agent
Decision memoryNever asked twice
Generation reuseCarried across runs
Governance
Access policiesUser, group, agent
Scoped tokensLeast privilege
Review & approvalAgents and definitions
Audit trailWho did what
Your environment Your computeWorkers run the agents
Agent runtimeYour code and framework
Model callsYour models and keys
Your data estateStays in place
DatabasesData catalogsAnalytics platformsDocumentsUser feedback
One loop. Feedback and gaps flow back into the engine.
The foundation

One semantic layer, built autonomously.

Colrows reconciles what your organisation already knows into one shared layer.

Sources
DatabasesSchemas, profiles, query history
Data catalogsOwners, lineage
Analytics platformsDashboards, BI models
DocumentationWikis, specs, PDFs
User feedbackCorrections, approvals
Semantic inference engine
  1. 1
    Extract signalsNames, usage, definitions, joins.
  2. 2
    Reconcile claimsAgreement builds confidence; conflicts surface.
  3. 3
    Propose definitionsEach backed by evidence.
  4. 4
    Curate and rememberDecisions are kept.
Semantic layer
Shared business meaning

One definition of every number, and where it comes from.

MetricsEntitiesBusiness termsRelationshipsVerified queriesAccess policies
Evidence-backedVersionedGoverned
Under the hood
Weighted signals

Every source becomes weighted evidence. Conflicts go to a curator.

schemasquery historyBI modelsdocumentscorrections
Reads your BI
TableauPower BILookerQlikThoughtSpotSuperset
13 governed object types

Versioned, and served to scoped clients.

Business termsMetricsDatasetsTablesColumnsTable relationshipsRelationshipsObjectsEventsActionsCategoriesCategory setsVerified queries
Grounded access

One meaning for every agent.

Data, documents and memory, through one layer and one set of access rules.

One governed interfaceREST and MCP, with the caller’s own access.
Scoped tokensAccess policiesAudited
Structured data

Live connectivity

Governed SQL, run live on your warehouses.

  • Each user’s own access
Unstructured data

Documents

Answers from wikis and PDFs, every claim cited.

  • Respects each user’s document permissions
Context over time

Persistent memory

Preferences and past decisions, kept across runs.

  • Private per user or agent
Semantic layerMetrics, entities, terms, verified queries.
Under the hood
19 tools, one catalogue
Semantic
searchSemanticEntitiesgetSemanticEntitiesgetJoinGraphresolveValueslistDatasourcessearchSqlResolutionsfindSqlResolutionMatch
SQL
validateSQLcompileSQLexecuteSQLgetSQLExecutionStatusgetSQLExecutionResultcancelSQLExecution
Documents
askDocuments
Memory
recallAgentMemoryrememberAgentMemoryforgetAgentMemory
Agent runs
getRunContextreportSemanticGap
MCP tool reference
Credentials and scopes
  • OAuth clientsServer-to-server
  • MCP appsChatGPT, Claude, Cursor and more
  • Agent run tokensPer run, approved tools only
semantic:readsemantic:read-writedata:querydocuments:readmemory:readmemory:writeaccess-control:read
Authentication and scopes
Agents at work

Governed agents, run on your compute.

Approved in Colrows, run on workers in your environment.

Agents
AI AnalystOut of the box

Answers with governed SQL and cited documents.

CuratorOut of the box

Resolves inquiries, proposals and reported gaps.

Your own agents

Same guardrails, any framework.

In Colrows
Agent control plane
  1. 1
    Register and versionOwner, spec, tools, scope.
  2. 2
    Review and approveApproved before it runs.
  3. 3
    TriggerOn demand or scheduled.
  4. 4
    Check every answerVerified against run evidence.
  5. 5
    AuditWho ran what, as whom.
In your environment
Your computeWorkers claim and execute runs.
Agent runtimeYour code and framework
Model callsYour models and keys
Run tokensShort-lived, never raw SQL
⇆
Tools call back to ColrowsOver REST or MCP
Under the hood
Declarative spec

What it may touch, how it is judged.

semanticScopetoolsidentitylimitstriggersoutput contractevaluation
Guard on every call

Checked against spec and budget.

TOOL_NOT_ALLOWEDOUT_OF_SCOPERAW_SQL_DENIEDBUDGET_EXCEEDED
Evidence check

Claims must point to this run’s own evidence.

query executionsdocument citations
Lease-based workers

Workers hold a lease on each run.

claimheartbeateventscompletefailyield
Agent control plane docs
The loop closesDecisions and reported gaps flow back into the semantic layer.Back to the foundation
Governance

Governance runs through every layer.

Enforced by the platform, not by each agent.

Access policies

Per user, group and agent, on every call.

row-levelcolumn-levelmasking

Scoped tokens

Short-lived, least privilege.

OAuthMCP appsrun tokens

Review & approval

Agents and definitions, approved first.

approveretireevaluations

Audit trail

Who did what, as whom.

runstool callsdecisions

Questions about the platform.

What is the agent control plane?

Where agents are governed: registered with an owner and scope, approved before they run, checked against evidence and audited.

Where do my agents run?

On your infrastructure, with your code and model keys. Each run gets a short-lived token limited to approved tools, never raw SQL.

Which agents are included?

An AI Analyst that answers with governed SQL and cited documents, and a Curator that keeps the semantic layer healthy. Your own agents follow the same rules.

What does the semantic inference engine do?

It turns your databases, catalogs, BI tools, documents and feedback into weighted evidence, proposes definitions, and sends conflicts to a curator.

Is Colrows still a semantic layer?

Yes. It is the foundation that people, applications and agents all use.

See it on your data.

A 30-minute walkthrough of the platform on your own data.