Every option accepted by Agentronics.init, the AgentronicsClient surface, and the env vars consumed by the gateway.

Configuration

The browser SDK is configured with a single Agentronics.init({...}) call. The gateway is configured with environment variables. Every option is documented below — anything not listed here is treated as forward-compatible noise and ignored.

`Agentronics.init(options)`

import { Agentronics } from '@agentronics/sdk'
 
const client = Agentronics.init({
  publishableKey: 'agtx_pk_live_…',
  siteId: 'shop',
  policies,            // from sync.pullConfig()
  toolDescriptions,    // from sync.pullConfig()
  trace: { webhookUrl: '/api/agentronics' },
})

Option	Type	Default	Description
`publishableKey`	`string`	required	Publishable key (`agtx_pk_`). Secret keys (`agtx_sk_`) throw `wrong_key_kind`.
`siteId`	`string`	`'default'`	Stable identifier for this site — propagated onto every trace and gateway request.
`gatewayUrl`	`string`	`https://api.agentronics.dev`	Local development only (`localhost` / `127.0.0.1`); the retired `https://gateway.agentronics.dev` maps to the default; anything else throws.
`compat`	`'0.6'`	—	Temporary: keep the 0.6 behaviour while upgrading. See Upgrading to 0.7.
`debug`	`boolean`	`false`	Enables the console trace exporter and verbose warnings.
`policies`	`PolicyRule[]`	`[]`	Authorization rules — usually `pullConfig().policies` from the console.
`toolDescriptions`	`Record<string, string>`	`{}`	Approved descriptions from the console's Memory page; replace the code's text when tools are published.
`review`	`Reviewer \| false`	built-in dialog	Who approves `review` decisions. A function `(request) => Promise<boolean>`, or `false` to refuse them all.
`autoDetect`	`boolean \| DetectAgentOptions`	`true`	Run detection on init and attribute tool calls to the detected agent.
`visionDetection`	`boolean \| VisionDetectorOptions`	on with `autoDetect`	Watch interactions for vision-model agents. Reads timing and coordinates only.
`siteMemory`	`Partial<SiteMemory>`	—	Provide site context up-front; equivalent to calling `client.provideSiteMemory(...)` post-init.
`auth`	`AuthContext`	—	`verifyToken(method, token)` — upgrade browser-presented tokens from `declared` once your backend verifies them.
`progression`	`ProgressionConfig`	—	Tool surfacing stages — see Tool Management.
`tools`	`GovernedTool[]`	`[]`	Register tools synchronously instead of via `client.registerTool()`.
`trace.enabled`	`boolean`	`true`	Master switch for the tracer.
`trace.bufferSize`	`number`	`200`	In-memory ring buffer size before backpressure kicks in.
`trace.traceRate`	`number`	`1.0`	Sampling rate for non-error traces (`0.0`–`1.0`).
`trace.errorRate`	`number`	`1.0`	Sampling rate for error traces.
`trace.scrubPaths`	`string[]`	`[]`	JSONPath-ish paths scrubbed before export.
`trace.webhookUrl`	`string`	—	Send traces to your relay route (`createTraceRelay`), which forwards them to the console.
`trace.gateway`	`boolean`	`false`	Send traces straight to the Agentronics API with the publishable key (the page's origin must be on the key).
`trace.exporters`	`TraceExporter[]`	`[]`	Custom exporters appended to the pipeline.

`AgentronicsClient`

init returns a frozen client. The full type lives in @agentronics/sdk.

client.detect()                        // → AgentIdentity | null
client.authenticate({ … })             // → AgentIdentity | null
client.presentIdentity({ … })          // → PresentedIdentity
client.evaluate('checkout')            // → PolicyEvaluation
client.setPolicies(rules)              // replace authorization rules
client.syncConfig()                    // pull rules + approved descriptions + site memory (publishable key)
client.ping()                          // check the key against the API
client.registerTool({ name, … })       // → off()
client.publishToWebMcp()               // → { surfaced, registrations: Promise<WebMcpRegistration[]> }
client.setToolDescriptions(map)        // apply newly approved descriptions
client.estimateTokens()                // → { total, perTool }
client.installDomEnforcer()            // → { uninstall() }
client.provideSiteMemory({ … })        // → SiteMemory
client.traces()                        // → TraceEvent[] (in-memory)
client.vision?.score()                 // vision-agent score, 0–1

Server entrypoints

Import	What
`@agentronics/sdk/server`	`createAgentAuth`, `createAgentAuthHandler`, `expressAgentAuth`, `readAgentHeaders`, `toTraceEvent`, `createTraceRelay`, `createAgentContextHandler`, `protectedResourceMetadata`, `oauthChallenge`, `generateAgentKey`, `hashAgentKey`, `staticKeyVerifier`, `PUBLISHED_IP_RANGES`
`@agentronics/sdk/next`	`agentronicsMiddleware(options)`
`@agentronics/sdk` (server use)	`createDashboardSync({ url, ingestKey, siteId })` → `pushTools`, `pushMemory`, `pullConfig`, `keyVerifier`; `createDashboardExporter({ url, ingestKey })`

Bundle entrypoints

Import	What it ships	Budget
`@agentronics/sdk`	Full SDK — every pillar wired through `init()`.	≤ 30 KB brotli (`init` graph ≤ 28 KB)
`@agentronics/sdk/lite`	WebMCP-only `init` plus `detectWebMcp`.	≤ 8 KB brotli
`@agentronics/react`	Provider + hooks. Peer-deps on the SDK.	≤ 2 KB brotli

The CI pnpm size run in packages/sdk enforces all three plus a tree-shake budget on the Agentronics.init reachable graph.

Configuration

Configuration

Agentronics.init(options)

AgentronicsClient

Server entrypoints

Bundle entrypoints

On this page

`Agentronics.init(options)`

`AgentronicsClient`