Skip to main content

The agent tool families

The Model Context Protocol surface is how an agent, ours or someone else's, reaches the platform. This page is what is on it.

The naming contract​

Every tool name is exactly two snake_case segments: family.verb.

^[a-z][a-z0-9_]*\.[a-z][a-z0-9_]*$

That is enforced at registration rather than documented as a convention. A name that does not match is rejected, and two different families claiming one name is a fail-loud boot error rather than a silent last-one-wins.

Re-registering the same name is idempotent, so a family loaded twice is harmless. Two families disagreeing about who owns a name is not.

The surface is whole, or the cell does not boot​

This is the property that makes any count of this surface trustworthy.

Tool families are registered at boot. A mandatory family that fails to import or register is treated as invisible capability loss: failures are aggregated and raised as a single error that aborts boot.

The set of families permitted to be absent is empty. Nothing is optional.

So a running cell either has the whole tool surface or is not running. A partial load that quietly shrinks what an agent can do, which is the failure mode that makes tool counts meaningless elsewhere, is not reachable here.

The families​

Grouped by what they reach. Names below are the public capability names; the tool prefix is the lowercase family segment.

Data platform​

FamilyReaches
Catalog (read and write)Catalog assets, two families so read and write authority are separable
DatabaseThe database engine surfaces
DataGeneral data operations
LakehouseIceberg and Delta tables
StructuresThe structure engine
OntologyObject, link and action types
Graph write-backWriting through the graph projection
SemanticThe semantic layer and federated semantics
IngestionBringing data in
Data PipelinesThe ETL fabric
FunctionsUser-defined functions
StorageObject storage
MessagingThe durable bus
Data ExchangeCross-organization listings
Clean roomsRoom-scoped operations
DiagramsGenerated diagrams over an estate

Agents and intelligence​

FamilyReaches
AgentsAgent lifecycle and operation
Agent TeamsCrews, formations and billets
Orchestration and orchestration templatesThe workflow graph
Scheduled Jobs orchestrationLeased, scheduled work
CognitionFederated cognition
KnowledgeKnowledge fusion and canonical entities
The MindThe workspace and node roll-up
Model OperationsThe ML control plane
Algorithm LibraryKnown methods an agent can reach for
CodingGoverned coding runs: submit, status, cancel
ComputeKernels and execution
CopilotThe assistant surface
ContractsEntity contracts
Governed shellThe permission-gated shell surface

Security and governance​

FamilyReaches
SIEMSecurity events
SOARResponse orchestration
UEBABehaviour analytics
Advanced securityThe deeper security surface
Cyber OperationsThe governed assessment surface
SurveillanceObservation to canonical entity
AuditThe audit planes
Activity LedgerThe per-workspace hash-chained record
SecretsThe secret substrate
Security provisioningProvisioning security material
DirectoryEntities, groups and roles
FederationExternal identity providers
SessionsSession administration
ScopeScope resolution
ElectionsEntity voting and quorum
Worker PolicyWhich workers run where

Estate and operations​

FamilyReaches
EstateEstate-as-code
SystemCore system operations
FleetEdge fleet and releases
Workspaces and workspace provisioningWorkspace lifecycle
MeteringConsumption and cost
Disaster recoveryDR policy, drills, failover
Discovery & MappingScanning an estate in place

Mission​

FamilyReaches
TacticalTactical operations
Tactical CommunicationsSpectrum and radio control plane

Every call is governed​

A tool being on the surface does not mean an agent can call it. Each invocation crosses:

  • The agent's own cryptographic allowed-tools ceiling, which is per leased sub-agent, so two agents in one run can hold genuinely different powers.
  • The permission gate, whose layers are identical on a cell and on a device. See chat on core and edge for the four layers and the rule that a permissive autonomy mode removes prompts, never possibility.
  • The decision gate, for consequential actions.

Schema validation runs on arguments, and a path-segment validator rejects traversal.

Authentication​

MechanismStatus
Session and bearer tokenShipping
OAuth 2.1 resource serverPresent and off by default. The configuration block is empty unless a deployment populates it

A token-bucket rate limiter sits in front of the surface.

Resources, not only tools​

Alongside tools, the server registers MCP resources, under the same fail-loud discipline: a mandatory resource family that fails to register aborts boot rather than silently reducing what an agent can see.

Enumerate the live set, do not trust a list​

Counts age, registries do not

This page groups the families so you know the shape of the surface. It is not the authoritative list of tools, and neither is any number written down elsewhere.

The surface is registered at runtime. The correct way to know what a given deployment exposes is to enumerate the registry on that deployment, which is also how the platform's own measured figures are produced. A count harvested from source rather than from a running registry has been wrong before.

What you can rely on structurally: registration is fail-loud and the optional set is empty, so whatever a running cell reports is the whole surface rather than a partial load.