StableMindIntegrationsSIO25 · Integration Studio
StableMind sits between AI agents and consequential enterprise action, so integration has to be more precise than “the API works.” This catalog shows the adapters, protocols, connector boundaries, credential infrastructure, SDKs and ecosystem machinery already present in the sealed engineering source, together with the rule that keeps them safe: reachability is capability, not permission.
The integration doctrine
StableMind separates connector capability from semantic authority. An API scope, MCP tool registration, tenant role, OAuth token, vault lease or network route can describe what software is technically capable of reaching. None of those facts answers whether this machine may perform this consequential action against this target now.
Authority Cloud resolves delegated power, ActionGate decides the semantic request, and an exact Action Permit constrains the allowed target, limits and time. Connector configuration therefore remains downstream of a decision boundary instead of becoming an accidental permission system.
A connector may produce an attributable execution receipt. Evidence can later compare expected postconditions with independently observed state. The fact that an integration returned HTTP 200, a tool result, or a vendor success code does not by itself prove the intended consequence.
Source-backed catalog
Every card below maps to source that already exists in the sealed ActionGate lineage. “Engineering ready” means there is source-backed engineering material and testable contracts. It does not mean a legal customer has connected the integration, that an external vendor has certified it, or that StableMind has exercised it in a live customer environment.
12 integration records visible.
Govern Model Context Protocol tool calls as semantic action requests before consequential tool execution. Exact tool, arguments, target and authority bindings remain subject to ActionGate policy.
Boundary: MCP connectivity does not grant tool authority. A reachable tool still requires current delegated authority and an exact decision.
Source-backed Microsoft enterprise action evaluation with exact profile and authority binding. Designed to govern semantic enterprise actions without equating tenant access, app registration or token possession with permission.
Boundary: Tenant access and Microsoft credentials are capability inputs, never delegated machine authority.
Evaluate Salesforce actions against exact tenant, operation, object and authority bindings before execution. The adapter preserves semantic-action boundaries instead of treating an OAuth scope as permission to perform every reachable mutation.
Boundary: OAuth scopes constrain technical reach; they do not manufacture authority for a consequential Salesforce action.
Govern ServiceNow enterprise operations with exact instance, domain, role and operation profiles. Source contracts reject encoded-query expansion, dot-walk drift, role supersets and session identifiers at the authority boundary.
Boundary: ServiceNow role membership or API reach does not authorize an agent to exercise every available operation.
A bounded connector-host contract for running certified connector processes with deterministic handshakes, request binding, deadlines and receipts. It is the runtime boundary between an exact permit and an implementation-specific integration.
Boundary: The host cannot broaden the Action Permit, invent a semantic action or retain credential material as standing capability.
Source-backed certification machinery checks connector packages, publisher signatures, subprocess behavior, credential redaction and deterministic receipts against the StableMind Connector Standard.
Boundary: A passing connector certificate proves package conformance to a scoped engineering contract. It never grants tenant-local action authority.
Go, Python and TypeScript SDKs implement actiongate.connector/1.0 framing, invocation binding, credential redaction and deterministic result receipts for custom connectors.
Boundary: SDK adoption makes integration easier; it does not bypass connector admission, authority checks, permits or customer policy.
Materialize short-lived technical capability only after an exact permit exists. Profiles bind route, target, credential type, scope, TTL and use count, with volatile handling and explicit revocation.
Boundary: Credential availability is downstream of authority. The broker cannot issue permission and cannot upgrade a commercial identity into machine authority.
A payment-specific connector contract with preflight, exact action binding, JIT credential handling and reconciliation semantics for consequential financial workflows.
Boundary: Payment rails and bank credentials remain execution capability. Destination, amount and authority must still be independently governed.
A source-backed supplier-information boundary emphasizing upstream identity, field minimization, freshness and a direct-mutation prohibition. Useful for authority and target resolution without turning read access into mutation power.
Boundary: The connector is read-oriented by contract and does not create supplier-change authority.
Signed extension manifests, publisher identity, SBOM/provenance digests and declared capabilities support a governed extension ecosystem for connectors and evidence verifiers.
Boundary: Gallery publication, publisher verification and extension installation create no standing authority and bundle no credentials.
An open lab certification bundle exercises authority-aware ecosystem interoperability and conformance without promoting synthetic lab success into third-party production proof.
Boundary: Open ecosystem verification is engineering evidence only. It cannot certify a customer workflow or create action authority.
What Integration Studio actually configures
OIDC identity can establish who is using the StableMind commercial workspace. That identity remains separate from machine delegation. SIO19 deliberately avoids embedding identity-provider secrets and production authentication still fails closed until a deployment configures a real provider.
SIO23 can prepare customer-controlled read-only observation paths. Integration Studio can classify a source for observation or evidence, but it does not silently convert a read connection into an execution connector or broaden the seven-field minimized shadow envelope.
Execution connectors sit behind permits. Credential Broker materializes short-lived capability only after permit eligibility exists. The connector SDKs make implementation predictable, while connector certification checks package conformance without granting tenant authority.
Six integration roles
The same vendor can appear in more than one place in an agentic architecture. A Microsoft or Salesforce boundary might be used only as a read-only observation source during shadow evaluation, then later as an execution connector for a narrow semantic action. StableMind keeps those roles explicit so a low-consequence connection cannot silently inherit the powers of a high-consequence one.
An identity integration establishes an attributable human or service identity for the commercial workspace and can support enterprise sign-in policy. It does not establish what an AI agent may do. That distinction prevents an authenticated administrator, group membership, SSO assertion, or tenant role from becoming an accidental delegation mechanism.
An observation source supplies minimized, read-only facts about a workflow so ActionGate can compute counterfactual decisions. Observation contracts should specify exact fields, freshness, tenancy, transport ownership and prohibited secret classes. Read access is deliberately narrower than execution, and a shadow source must not acquire mutation capability merely because the integration supports it.
An agent runtime integration identifies the semantic request coming from an agent framework, tool protocol, orchestration system or custom application. StableMind cares about the meaning of the requested action, the consequential target and the authority lineage behind it. Runtime identity or tool registration alone cannot answer those questions.
An execution connector implements the final system-specific operation only after a valid permit exists. The connector should bind the permit, semantic action, target, request digest, deadline and idempotency state; receive technical capability only when required; and emit an attributable receipt without broadening the decision made upstream.
A credential provider or broker supplies the smallest short-lived technical credential required by the exact permitted action. It should support tight TTLs, bounded use counts, volatile handling and revocation. StableMind treats that credential as capability, not permission, which means no vault record or cloud role can substitute for ActionGate authorization.
An evidence destination receives receipts, verifier observations, lineage or other governed artifacts under customer-controlled custody. It does not decide whether the action was authorized and it does not turn a successful connector response into Proof of Consequence. Evidence remains useful precisely because it stays distinct from both the actor and the decision boundary.
Evaluation checklist
What exact semantic action can the integration perform? What consequential target does it affect? Which explicit delegation proves that the machine may exercise that power rather than merely reach the API?
When does technical capability appear, how short-lived is it, and can it be revoked? Which consequence dimensions need reserved capacity before the action is eligible to proceed?
What receipt can the integration produce, what independent observation can verify the resulting state, and which Guardian path can interrupt or revoke power before the consequence completes?
Enterprise evaluation
Use the source-backed Enterprise Evaluation Room to review security, architecture, deployment, procurement, evidence and pilot design without turning evaluation completeness into production proof.
Open Enterprise Evaluation