Agent service runtime

An agent service runs your agent as its own process, independent of the app server. Use it when you need a separate process boundary, direct control-plane registration, remote MCP tools, or deployment-owned service telemetry. Use a normal in-app route for everything else. Veryfront Cloud can invoke a push runtime directly against an agent service, which is the main reason to deploy one even when the app and the agent share a host. Shared and managed dedicated servers use the framework-owned veryfront serve runtime instead. That runtime discovers all project agents and tools, then routes each signed control-plane request by agentId. Projects on a managed dedicated server do not require a service.ts entrypoint. Add one only when you intentionally run the standalone Agent Service process described in this guide.

Prerequisites

At least one agent in agents/ that the service should expose (see Agents).
A deployment target you can run a long-running Node process on.
For Veryfront Cloud registration: VERYFRONT_API_TOKEN, VERYFRONT_PROJECT_ID or VERYFRONT_PROJECT_SLUG, and a publicly reachable VERYFRONT_AGENT_SERVICE_URL. See Configuration for the full list.
Immutable deployment metadata for runtimeSource when the control plane invokes the service.

Create a service entrypoint

Create a process entrypoint that starts the default Veryfront Cloud agent service runtime:

// service.ts
import { startNodeVeryfrontCloudAgentService } from "veryfront/agent";

await startNodeVeryfrontCloudAgentService();

startNodeVeryfrontCloudAgentService() starts the runtime from the environment that is already loaded. It does not load local .env files or initialize process-wide telemetry. Load standalone service environment files through the trusted deployment wrapper before it imports service.ts. Project modules cannot mutate the shared process environment through the public agent API. Initialize service-level OpenTelemetry in the trusted deployment wrapper before it loads service.ts. Do not let project code select process-wide exporters, trace hooks, or application-error reporters. The framework-owned veryfront serve runtime owns this setup on shared and managed dedicated servers. The service discovers the same project primitives as the app runtime:

agents/
tools/
skills/
resources/
prompts/
workflows/
tasks/

When exactly one code or markdown agent is discovered, that agent becomes the default for direct /api/runs requests. Pass agentId when the service exposes multiple agents and direct requests need a predictable default.

Keep agent behavior in project files

Define the agent in agents/ and keep service startup separate from agent behavior:

// agents/support.ts
import { agent } from "veryfront/agent";

export default agent({
  id: "support",
  system: "You help users resolve support issues.",
  tools: {
    searchDocs: true,
  },
});

Markdown agents use the file path as the agent id:

---
name: Support
description: Helps users resolve support issues
max-steps: 6
---

You help users resolve support issues. Ask for missing details before acting.

For non-standard project layouts, configure discovery paths in veryfront.config.ts under ai.<primitive>.discovery.paths.

Configure registration

Control-plane registration is convention-first. In auto mode, the service registers only when VERYFRONT_API_TOKEN and VERYFRONT_AGENT_SERVICE_URL are present. Registration also requires the immutable runtimeSource binding described below.

VERYFRONT_API_URL=https://api.example.com
VERYFRONT_API_TOKEN=<TOKEN>
VERYFRONT_PROJECT_ID=<PROJECT_ID>
VERYFRONT_AGENT_SERVICE_URL=https://agent.example.com
VERYFRONT_AGENT_SERVICE_REGISTRATION=auto

Use VERYFRONT_AGENT_SERVICE_REGISTRATION=enabled when startup must fail if the service cannot register. Use disabled when the service must run without control-plane registration. The service name resolves from VERYFRONT_AGENT_SERVICE_NAME, then the nearest package.json or deno.json name, then veryfront-agent-service. Pass serviceName only when code should override that convention.

Bind control-plane runs to the deployed source

A standalone agent service discovers one local project snapshot at startup. It cannot select another project branch or release for an individual request. Bind the service to deployment-owned immutable metadata when it accepts signed control-plane runtime invocations:

import { startNodeVeryfrontCloudAgentService } from "veryfront/agent";

const environmentName = process.env.DEPLOYED_ENVIRONMENT_NAME;
const releaseId = process.env.DEPLOYED_RELEASE_ID;
if (!environmentName || !releaseId) {
  throw new Error("Missing immutable agent service deployment identity");
}

await startNodeVeryfrontCloudAgentService({
  runtimeSource: {
    type: "environment",
    environmentName,
    releaseId,
  },
});

The service accepts a control-plane invocation only when its agentSource exactly matches runtimeSource. An unbound service returns CONTROL_PLANE_AGENT_SOURCE_UNBOUND. A different release or environment returns CONTROL_PLANE_AGENT_SOURCE_MISMATCH. Branch sources are mutable and return CONTROL_PLANE_AGENT_SOURCE_UNSUPPORTED. Do not resolve runtimeSource from the latest deployment at request time. Pass the environment and release identifiers that produced the running service artifact. Direct /api/runs requests do not select project source and do not require this binding.

Add remote MCP tools

Use mcpServers when the service needs remote tools. Use veryfrontApiMcpServer() and veryfrontStudioMcpServer() for Veryfront-owned control-plane MCP servers and normal MCP server config objects for third-party servers. This service startup config uses endpoint and headers. Per-agent config in agent({ mcpServers }) uses transport, auth, and toolPolicy.

import {
  startNodeVeryfrontCloudAgentService,
  veryfrontApiMcpServer,
  veryfrontStudioMcpServer,
} from "veryfront/agent";

await startNodeVeryfrontCloudAgentService({
  serviceName: "support-agent",
  mcpServers: [
    veryfrontApiMcpServer(),
    veryfrontStudioMcpServer(),
    {
      id: "linear",
      endpoint: process.env.LINEAR_MCP_URL,
      headers: {
        Authorization: ["Bearer", "<TOKEN>"].join(" "),
      },
      toolPolicy: {
        allow: ["search_issues", "create_issue"],
        approval: "never",
      },
    },
  ],
});

If mcpServers is omitted, the Veryfront Cloud preset includes veryfrontApiMcpServer() by default. Pass mcpServers: [] to run without remote MCP tools.

Reach trusted deployment-local MCP servers

The default remote MCP source uses guarded outbound networking. Keep that default for third-party, request-derived, and tenant-configured endpoints. A separately deployed agent service may need to reach a trusted MCP server on a private cluster address. In that case, capture the host transport and the exact allowed endpoints once at startup. Use the host transport only for those immutable endpoints and preserve the guarded source for everything else:

import {
  startNodeVeryfrontCloudAgentService,
  veryfrontApiMcpServer,
  veryfrontStudioMcpServer,
} from "veryfront/agent";
import { createRemoteMCPToolSourceFactoryWithTransport } from "veryfront/tool";

function requiredUrl(name: string): string {
  const value = process.env[name];
  if (!value) throw new Error(`Missing ${name}`);
  return value;
}

const hostFetch = globalThis.fetch.bind(globalThis);
const createRemoteToolSource = createRemoteMCPToolSourceFactoryWithTransport({
  trustedEndpoints: [
    requiredUrl("VERYFRONT_MCP_URL"),
    requiredUrl("VERYFRONT_STUDIO_MCP_URL"),
  ],
  requestFetch: hostFetch,
});

await startNodeVeryfrontCloudAgentService({
  createRemoteToolSource,
  mcpServers: [
    veryfrontApiMcpServer(),
    veryfrontStudioMcpServer(),
  ],
});

The framework rejects invalid allowlist entries at startup and uses the host transport only for an exact normalized URL match. Unmatched, invalid, and resolver-based endpoints retain guarded outbound networking. http: is appropriate only for private deployment-local networking; use https: for public networks. Never put a callback endpoint or a per-request URL in the trusted endpoint list.

Refresh runtime state

Use resolveRuntimeState when a long-lived service run must refresh instructions, context, or available tools at a model step boundary.

import { agent } from "veryfront/agent";

export default agent({
  id: "support",
  system: "You are a support assistant.",
  resolveRuntimeState: async ({ step }) => {
    if (step === 0) return;

    return {
      system: "Use the latest project instructions and tool inventory.",
    };
  },
});

Services that use Veryfront Cloud project steering can reuse fetchDefaultAgentServiceProjectSteering() for the initial fetch and createDefaultAgentServiceProjectSteeringRefresh() for step-boundary refresh.

Keep inference authority separate

Signed runtime invocations may include an optional credentials.inferenceAuthToken alongside the broader credentials.authToken. The inference credential is bound to the exact run and agent and is intended only for attributed Veryfront Cloud model requests. It is size-bounded to 16 KB, uses visible ASCII token characters, and remains optional so existing producers and consumers stay compatible. Treat this as trusted-host authority. Do not copy it into project context, tools, logs, durable request payloads, or a general API client. Framework-managed agent services bind it only after the signed invocation and run-event credential have been verified, bypass project model overrides for Veryfront Cloud models, and send it only to an HTTPS (or loopback development) gateway. Custom runtime adapters should follow the same boundary: use authToken for project and control-plane operations, and expose inferenceAuthToken only to the model subprocess’s Veryfront Cloud gateway configuration.

Use lower-level helpers

Use startNodeVeryfrontCloudAgentService() for the standard service shape. Use lower-level helpers only when the service needs a custom server adapter, custom execution preparation, or custom infrastructure.

Helper	Use
`defineAgentService()`	Normalize one or more agents into a service registry contract.
`startNodeAgentService()`	Start a Node service around a request-native runtime.
`prepareVeryfrontCloudAgentServiceChatExecution()`	Prepare Veryfront Cloud chat execution with model, steering, and durable-run wiring.
`createAgentServiceProjectSteering()`	Bind markdown agent definitions to project steering and skill refresh.

Migrate custom durable child event writers

This migration applies to custom hosted runtimes that call the lower-level durable child helpers. Framework-managed startNodeVeryfrontCloudAgentService() runtimes create and scope writer capabilities internally. Raw authToken, apiUrl, and runEventAppendToken fields no longer grant durable child event-writer authority. The parsed hosted request also excludes the writer credential. Keep the credential inside trusted ingress and replace the removed fields with an opaque HostedRunEventWriterCapability. Apply the change at every integration point your custom runtime implements:

Integration point	Migration action
`ParsedHostedChatRequest` / `ParsedAgentServiceChatRequest`	Stop reading `runEventAppendToken`; verified ingress retains it privately.
`PrepareHostedConversationRootRunContextInput` / `PrepareAgentServiceConversationRootRunContextInput`	Remove `runEventAppendToken`; keep the exact-root capability in trusted host composition.
`ExecuteHostedDurableChildForkInput`	Pass the exact-parent capability; this helper mints the exact-child capability after persistence.
`DefaultHostedInvokeAgentToolOptions`	Pass the current run’s exact-parent capability.
`ExecuteHostedChildForkWithPreparedToolsInput` / `ExecuteHostedChildForkToolInputOptions`	Pass a capability bound to `durableChildRun.childRunId`.
`HostedDurableChildForkRunContextInput`	Remove `authToken` and `apiUrl`; pass the exact-child capability.
`HostedDurableRunStartExecutionInput`	Accept the required application-facing `rawRequest` in the starter callback.
`HostedAgentServiceDetachedExecutionInput` / `AgentServiceDetachedExecutionInput`	Accept the required application-facing `rawRequest`; internal control headers are already removed.

The generated veryfront/agent reference lists the complete properties for these contracts.

After trusted ingress verifies an exact root-run append credential, create the root capability. Do not pass a general user API token.

import {
  createHostedRunEventWriterCapability,
  executeHostedChildForkWithPreparedTools,
  executeHostedDurableChildFork,
} from "veryfront/agent";

const rootWriter = createHostedRunEventWriterCapability({
  apiUrl,
  runId: durableRootRun.runId,
  runEventAppendToken: verifiedRunEventAppendToken,
});

Pass that exact-parent capability to helpers that own child persistence and capability delegation. Do not pre-mint for these helpers.
const result = await executeHostedDurableChildFork({ ...input, runEventWriterCapability: rootWriter, });

For lower-level helpers that receive an already-persisted durableChildRun, mint and pass an exact-child capability:

const childWriter = await rootWriter.mintChildRunEventWriterCapability(
  durableChildRun.childRunId,
  abortSignal,
);

const result = await executeHostedChildForkWithPreparedTools({
  ...input,
  durableChildRun,
  runEventWriterCapability: childWriter,
});

Update detached starter callbacks to accept the isolated request:

const startDetachedExecution = async ({
  execution,
  abortSignal,
  rawRequest,
}: HostedAgentServiceDetachedExecutionInput<Execution>) => {
  await host.start({ execution, abortSignal, request: rawRequest });
};

A durable execution without authority bound to the expected run fails before provider dispatch. Token exchange failures are bounded, sanitized, and fail closed; callers must not retry by falling back to a user API token.

Verify it worked

Start the service entrypoint and call the run route directly. The default port is 3001; override with PORT if needed.

node service.ts &
curl -N http://localhost:3001/api/runs \
  -H "Content-Type: application/json" \
  -d '{"agentId":"support","messages":[{"id":"1","role":"user","parts":[{"type":"text","text":"ping"}]}]}'

A working service streams AG-UI events back. If Veryfront Cloud registration is enabled, the service should also appear in the cloud dashboard’s agent service list after the first heartbeat (VERYFRONT_AGENT_SERVICE_HEARTBEAT_INTERVAL_MS).

Getting Started

Concepts

Guides

API Reference

Agent service runtime

Prerequisites

Create a service entrypoint

Keep agent behavior in project files

Configure registration

Bind control-plane runs to the deployed source

Add remote MCP tools

Reach trusted deployment-local MCP servers

Refresh runtime state

Keep inference authority separate

Use lower-level helpers

Migrate custom durable child event writers

Verify it worked

​Prerequisites

​Create a service entrypoint

​Keep agent behavior in project files

​Configure registration

​Bind control-plane runs to the deployed source

​Add remote MCP tools

​Reach trusted deployment-local MCP servers

​Refresh runtime state

​Keep inference authority separate

​Use lower-level helpers

​Migrate custom durable child event writers

​Verify it worked

Prerequisites

Create a service entrypoint

Keep agent behavior in project files

Configure registration

Bind control-plane runs to the deployed source

Add remote MCP tools

Reach trusted deployment-local MCP servers

Refresh runtime state

Keep inference authority separate

Use lower-level helpers

Migrate custom durable child event writers

Verify it worked