Governed execution for AI agents

Let AI agents act. Without letting them loose.

TrustRail sits between your AI agents and the real world. It checks every consequential action against your rules, asks a human when it matters, and keeps your credentials out of the agent’s hands — with a tamper-evident record of everything.

Scans locally No Docker or terminal Bring your existing agents
TrustRail — Approvals
Awaiting approvalpayments.purchase.create$49.99 · order-4821digest 9f2c…a71b
Executedcommunications.email.sendsupport@example.comat most once
Deniedrepository.mergepolicy: protected-branchno token issued
Containedagent: release-bottier C4 · credentials revokedrunbook 4/5
everything is denied until you allow it ·approvals match the exact request ·agents never hold your keys ·risky actions wait for a human ·every action leaves a receipt ·nothing runs twice by accident ·

Why you need this

AI agents are genuinely useful right up until one buys the wrong thing, emails the wrong person, or deletes something that mattered. TrustRail exists for that moment.

The problem

Agents act faster than you can watch

An agent can send a hundred emails or place an order in the time it takes to glance at a screen. Hoping it behaves is not a plan.

The fix

One place to say yes, no, or ask me

You set simple rules once. TrustRail applies them to every action, every time — and anything risky waits for a tap from you before it happens.

The payoff

Proof of everything, after the fact

Every decision is recorded in a log that cannot be quietly edited. When someone asks “what did the agent actually do?”, you have the receipt.

How it works

Think of it as a seatbelt for your AI agents: they keep driving, and TrustRail decides what actually happens when they reach for the real world.

Connect your agentsThe desktop app finds the agents already on your computer, or you add one line of code with the SDK. No rewiring, no migration.
Set the rulesDecide what runs on its own, what gets blocked, and what waits for a person — like “any purchase over $50 needs my approval.”
TrustRail decidesSafe actions run instantly. Risky ones pause until you tap Approve. Everything is recorded so you can always see what your agents did, and why.

Bring your agents over. Keep the guarantees.

Everything below ships in the product today. Connect once, decide deterministically, and contain from one console when something goes wrong.

Setup assistant

Moving an agent in takes one pass

The desktop app scans this machine the way Migration Assistant scans an old Mac: it finds your MCP servers, agent configs and provider keys, shows you exactly what it found, and imports only what you tick.

It runs as part of the desktop app — no Docker, command line, or local stack is required to discover and prepare your existing agent setup.

It only offers providers this build can actually govern — the list comes from the action registry, not from a table that can drift. Credential values never leave the app; you enter the key in the console, where it is tested against the real provider and handed straight to the gateway.

Setup Assistant — Choose what to bring over
FoundClaude Code3 MCP serversReady
FoundStripeSTRIPE_API_KEYValid
FoundSlackSLACK_BOT_TOKENEnforce ready
Left aloneANTHROPIC_API_KEYmodel key, not an actionSkipped

Connection wizard

A credential is proven before it is stored

Every connection is tested with a real read-only call to the provider over the same pinned-origin transport the gateway uses. You get back one of three words — valid, invalid, unavailable — and never the provider’s response.

Connect stays locked until the current credential, origin, provider and environment have all produced a valid result. Change any one of them and the proof is void.

Connect a provider
StripeGET /v1/accountValid
Origin pinned to api.stripe.comEnforced
Credential stored in the gateway onlyNever returned

Incident runbook

Containment is a button, not a paragraph

Acknowledge, contain, revoke credentials, inspect the evidence, and recover — each one an executable control in the console, each one gated by the permission it actually needs and each one appended to the tamper-evident chain.

A freeze does not burn credentials, because the next rung has to have somewhere to escalate to. The console says which rung you are on and what it did.

Incident — release-bot
DoneAcknowledgeoperator@example.com
DoneContaintier C4
DoneRevoke credentialsunusable during recovery
NextInspect evidence41 ledger entries

The action lifecycle, governed end to end

An agent never holds provider credentials and never calls a provider directly. Every consequential action flows through five governed stages, each producing verifiable evidence.

Canonical actionThe agent submits a typed, normalized action. TrustRail computes a canonical SHA-256 digest that every later stage binds to.
Deterministic decisionA bounded, non-Turing-complete policy engine plus integer-only risk scoring decide allow, deny, or require-approval — bit-identically reproducible.
Human approvalApprovers confirm the exact subject digest with phishing-resistant MFA. Quorums, separation of duties, and race-safe single finalization built in.
Signed authorizationA five-minute Ed25519 token with a closed claim set and a one-use identifier is the only thing that unlocks execution.
Enforced executionA separate gateway process holds the provider credential, revalidates every fact, and treats ambiguous outcomes as first-class state to reconcile — never to retry blindly.

Built for the failure cases

Most agent frameworks optimize the happy path. TrustRail is engineered around what goes wrong: prompt injection, replayed requests, stale authority, crashed processes, and ambiguous provider timeouts.

Invariant

A denied action cannot execute

The gateway consumes a signed, digest-bound, one-use token before any provider call. No token, no execution — enforced in a separate process with its own least-privilege database role.

Invariant

Mutation invalidates authorization

Approvals bind to the canonical digest of the exact request. Change one byte of the payload and every downstream authorization fails closed.

Invariant

Delegation only narrows

Agent-to-subagent handoffs are signed grants checked on ten axes — capabilities, resources, spend, depth, time — in both the application and the database. Revoking an ancestor kills the whole subtree.

Invariant

Ambiguity is never a retry

A provider timeout is recorded as UNKNOWN and reconciled by a stable idempotency key. The guarantee is at most one logical operation — stated honestly, engineered thoroughly.

Invariant

Tenants are isolated in the database

Forced PostgreSQL row-level security with composite tenant keys means isolation holds even if application code has a bug. Cross-tenant probes return 404, never a hint.

Invariant

Evidence is tamper-evident

Every privileged transition appends to a per-tenant SHA-256 hash chain in the same transaction. The whole chain is re-verifiable through the API on demand.

Honest enforcement modes

TrustRail tells you exactly what each mode guarantees — and what it does not. Only Enforce mode prevents external action; the others are for visibility and rollout.

ModeBehaviorSecurity claim
ObserveReceives telemetry alongside or after actionsVisibility only — no prevention claim
ShadowEvaluates before the action; the client still executes independentlyPolicy impact measurement — no enforcement claim
DecisionReturns allow/deny/approval, but a client could bypass itAdvisory authorization
EnforceThe gateway holds the provider credential and the execution pathProtected against ordinary client bypass within the deployment boundary

Two integration modes, two different promises

Start by gating the tools you already have. Move the ones that matter inside the boundary. The difference is not convenience — it is what survives a compromised agent process, and TrustRail will not let you confuse the two.

Decision mode

guard() — one wrapper, no migration

Your tool keeps its signature and its implementation. TrustRail decides whether it runs, and a denial reaches the model as a tool result that tells it not to route around the refusal. Works with every action type in the registry.

Gates your code; does not hold your credentials. A compromised agent process could bypass it. Not enforcement.

Enforce mode

act() — the agent never holds the credential

Drop the local implementation. A separate gateway process holds the provider credential, revalidates every fact, and spends a one-use Ed25519 authorization. The agent cannot perform the action on its own.

Requires a reviewed provider adapter, so it covers the action types the registry marks EXECUTABLE.

// Enforce mode: evaluate, wait for a human if required, execute
const result = await agent.act({
  purpose: "Reorder the stock the planner flagged",
  actionType: "payments.purchase.create",
  resource: { type: "payment", id: "order-4821" },
  parameters: restockOrder,
  financial: { amountMinor: "4999", currency: "USD" },
  waitForApprovalSeconds: 120,
});

// -> result.status: "EXECUTED" | "DENIED" | "AWAITING_APPROVAL"

A denial is a returned result, never a thrown error. Refusing an action is a normal answer, and code that treats it as a crash ends up working around it.

Drops into the frameworks you already use

The adapters are matched structurally and import nothing, so there are no peer dependencies and no version coupling. Governance refusals become tool results rather than exceptions that abort a run.

FrameworkIntegration
Vercel AI SDKgovernedAiTool(agent, spec, tool)
LangChain / LangGraphgovernedLangChainTool(...) · Python @agent.guard(...)
OpenAI / Anthropic tool loopsgovernedHandler(agent, spec, handler)
MCP-capable agentsOne config entry — Claude Code, Claude Desktop, and others
Anything elsegovernedExecute · governedTextExecute