Official JavaScript/TypeScript SDK for Symbiont, the policy-governed agent runtime. Same agent. Secure runtime.
This SDK is the integration surface for the Symbiont runtime. Use it from Node.js, edge runtimes, or the browser to manage agents and workflows, drive scheduled and channel-bound execution, manage policies, secrets, and MCP connections, verify webhooks, and integrate AgentPin identity — all against a runtime that enforces Cedar policy, SchemaPin tool verification, ToolClad tool contracts, and tamper-evident audit logging.
The runtime decides what an agent may do. The SDK decides how your application talks to the runtime.
AI agents are easy to demo and hard to trust. Once an agent can call tools, access files, send messages, or invoke external services, you need more than prompts and glue code. You need:
- Policy enforcement for what an agent may do — built-in DSL and Cedar authorization
- Tool verification so execution is not blind trust — SchemaPin cryptographic verification of MCP tools
- Tool contracts for how tools execute — ToolClad declarative input validation, scope enforcement, and injection prevention
- Agent identity so you know who is acting — AgentPin domain-anchored ES256 identity
- Audit trails for what happened and why — cryptographically tamper-evident logs
- Approval gates for sensitive actions — human review before execution when policy requires it
Symbiont is the runtime that enforces all of this. This SDK is the typed, ergonomic way to drive it from JavaScript and TypeScript.
A running Symbiont runtime is required. The fastest way:
# Start the runtime (API on :8080, HTTP input on :8081)
docker run --rm -p 8080:8080 -p 8081:8081 ghcr.io/thirdkeyai/symbi:latest upFor Homebrew, install scripts, building from source, or production deployment, see the getting-started guide.
npm install symbi-coreimport { SymbiontClient } from 'symbi-core';
const client = new SymbiontClient({
apiKey: process.env.SYMBIONT_API_KEY,
apiUrl: 'http://localhost:8080/api/v1',
});
await client.connect();
// Create an agent
const agent = await client.agents.createAgent({
name: 'textProcessor',
description: 'Processes and analyzes text input',
parameters: [{ name: 'text', type: { name: 'string' }, required: true }],
returnType: { name: 'string' },
capabilities: ['text_processing'],
});
// Execute it
const result = await client.agents.executeAgent(agent.id, { text: 'Hello, Symbiont!' });
console.log(result.result);The SDK talks to the runtime over its versioned REST API (/api/v1).
SDK (symbi-core) |
Symbiont runtime |
|---|---|
| 1.14.x | 1.14.x (tested); compatible with 1.15.x / 1.16.x for the documented surface |
Pin the SDK's minor version to your runtime's minor version when you can; newer 1.x runtimes remain compatible for the endpoints documented here.
The SDK exposes one client (SymbiontClient) with namespaced sub-clients for each runtime feature.
| Sub-client | Surface |
|---|---|
client.agents |
Agent lifecycle (create, update, execute, re-execute, delete, history, heartbeat, push events) |
client.schedules |
Cron schedules with pause/resume/trigger and run history |
client.channels |
Slack/Teams/Mattermost adapters, mappings, audit |
client.workflows |
Multi-agent workflow execution |
client.agentpin |
Client-side AgentPin keygen, credential issuance and verification, discovery, key pinning |
client.policies / client.policyBuilder |
Policy creation and management |
client.secrets |
Vault, encrypted-file, and OS-keychain secrets |
client.toolReview |
Tool review workflow |
client.mcp |
MCP server connection management |
client.system |
Health, metrics, runtime info |
client.metricsClient |
OTLP / file metrics export, snapshots, periodic collection |
client.http |
Dynamic HTTP endpoint management |
Standalone modules (no client required): WebhookVerifier (HMAC, JWT, provider presets), MarkdownMemoryStore (file-based agent context), SkillScanner / SkillLoader (ClawHavoc scanning + skill loading).
The SDK gives you typed access to the Trust Stack primitives it exposes, against a runtime that enforces the rest:
- Cedar policies via
client.policiesand theclient.policyBuildernamespace — create and manage fine-grained authorization for agent actions - AgentPin via
client.agentpin.*— domain-anchored ES256 credential issuance and verification, runs entirely client-side (no runtime required) - Secrets via
client.secrets— Vault, encrypted-file, and OS-keychain backends - MCP connections via
client.mcpand tool review viaclient.toolReview - ToolClad tool contracts and SchemaPin verification — enforced by the runtime server-side; tool signatures are verified before tool execution
Model output is never treated as execution authority. The runtime controls what actually happens.
The SDK is published as a set of focused packages. Most users only need symbi-core (which re-exports types from symbi-types and re-exports the agent client).
| Package | Purpose |
|---|---|
symbi-core |
Main client, configuration, auth, webhook verification, skills, metrics, memory |
symbi-types |
Shared TypeScript interfaces and Zod schemas |
symbi-agent |
Agent, schedule, channel, workflow, AgentPin clients |
symbi-policy |
Policy builder and enforcement |
symbi-secrets |
Vault / file / OS-keychain secrets backends |
symbi-mcp |
MCP protocol client |
symbi-tool-review |
Tool review workflow |
symbi-testing |
Mocks and test helpers |
symbi-cli |
Command-line tooling |
symbiont-sdk-js |
Monorepo source distribution |
await client.connect();
// Create and execute an agent
const agent = await client.agents.createAgent({
name: 'textProcessor',
description: 'Processes and analyzes text input',
parameters: [{ name: 'text', type: { name: 'string' }, required: true }],
returnType: { name: 'string' },
capabilities: ['text_processing'],
});
const result = await client.agents.executeAgent(agent.id, { text: 'Hello, Symbiont!' });
console.log(result.result);
// List the fleet
const agents = await client.agents.listAgents();AgentPin operations run client-side; no Symbiont runtime is required.
const { privateKeyPem, publicKeyPem } = client.agentpin.generateKeyPair();
const kid = client.agentpin.generateKeyId(publicKeyPem);
const jwt = client.agentpin.issueCredential({
privateKeyPem,
kid,
issuer: 'example.com',
agentId: 'data-analyzer',
capabilities: ['read:data', 'write:reports'],
ttlSecs: 3600,
});
// Verify (fetches discovery automatically)
const result = await client.agentpin.verifyCredential(jwt);
console.log(result.valid, result.agent_id, result.capabilities);The runtime's HTTP Input handler dispatches webhooks to a running agent on the communication bus, or falls back to an on-demand LLM ORGA loop against ToolClad manifests when the agent is not running. The SDK ships typed responses for both shapes:
import type {
WebhookInvocationRequest,
WebhookInvocationResponse,
WebhookCompletedResponse,
WebhookExecutionStartedResponse,
} from 'symbi-types';
// Send to your runtime's HTTP Input endpoint, then parse:
function handle(resp: WebhookInvocationResponse) {
if (resp.status === 'execution_started') {
console.log('dispatched message', resp.message_id, 'in', resp.latency_ms, 'ms');
} else {
console.log('LLM completed via', resp.provider, resp.model);
for (const run of resp.tool_runs) {
console.log(' -', run.tool, run.output_preview);
}
}
}import { HmacVerifier, JwtVerifier, createProviderVerifier } from 'symbi-core';
// Provider preset (GitHub, Stripe, Slack, custom)
const verifier = createProviderVerifier('GITHUB', Buffer.from(secret));
verifier.verify(request.headers, Buffer.from(request.body));
// Manual HMAC with prefix stripping
const hmac = new HmacVerifier(Buffer.from(secret), 'X-Hub-Signature-256', 'sha256=');
hmac.verify(headers, body);import { MarkdownMemoryStore } from 'symbi-core';
const store = new MarkdownMemoryStore('/data/memory', 30);
await store.saveContext('agent-1', {
agentId: 'agent-1',
facts: ['User prefers dark mode'],
procedures: ['Always greet by name'],
learnedPatterns: ['Responds better to bullet points'],
metadata: { lastSession: '2026-02-15' },
});
const context = await store.loadContext('agent-1');import { SkillScanner, SkillLoader } from 'symbi-core';
// 10 built-in ClawHavoc rules (pipe-to-shell, eval+fetch, base64-decode-exec, etc.)
const scanner = new SkillScanner();
const findings = scanner.scanContent(content, 'SKILL.md');
const loader = new SkillLoader({
loadPaths: ['/skills/verified', '/skills/community'],
requireSigned: false,
scanEnabled: true,
});
const skills = await loader.loadAll();SYMBIONT_API_KEY=... # required
SYMBIONT_API_URL=http://localhost:8080/api/v1
SYMBIONT_ENVIRONMENT=productionconst client = new SymbiontClient({
apiKey: process.env.SYMBIONT_API_KEY,
apiUrl: process.env.SYMBIONT_API_URL,
environment: 'production',
validationMode: 'strict',
timeout: 30000,
});See the API reference for the full configuration surface.
The Symbiont runtime itself lives at thirdkeyai/symbiont.
Workspace packages renamed from the previously-unpublished @symbi/* scope to unscoped symbi-* names so they can actually be installed from npm. Packaging-only release; no API changes.
WebhookInvocationResponsediscriminated union covering bothexecution_started(runtime communication-bus dispatch) andcompleted(on-demand LLM ORGA loop) shapes from the Symbiont v1.10.0 HTTP Input handlerWebhookToolRun,WebhookInvocationRequest,WebhookInvocationStatuswith Zod schemas- ToolClad v0.4.0 backend strings (
http,mcp,session,browser) accepted onToolManifestInfo.backend - All packages aligned to Symbiont runtime v1.10.0
See CHANGELOG.md for the full history.
Apache 2.0. See LICENSE.
The SDK is part of the Symbiont project's Community Edition. For Enterprise licensing of the Symbiont runtime (advanced sandbox backends, compliance audit exports, AI-powered tool review, encrypted multi-agent collaboration, monitoring dashboards, dedicated support), contact ThirdKey.

