@tame-ai/sdk

TypeScript SDK for protecting AI agent tool calls with TAME runtime policies.

Install

pnpm add @tame-ai/sdk

Configure

TAME_BASE_URL=https://tameapp.vercel.app
TAME_API_KEY=tame_sk_replace_me
TAME_ENVIRONMENT=production

Usage

import { TameClient } from "@tame-ai/sdk";

const tame = new TameClient({
  baseUrl: process.env.TAME_BASE_URL!,
  apiKey: process.env.TAME_API_KEY!,
  agentId: "code-pipeline-agent",
  context: {
    environment: process.env.TAME_ENVIRONMENT ?? "production",
  },
  failureMode: "fail_closed",
});

const applyPatch = tame.protectTool({
  name: "apply_patch",
  async execute(args: {
    repository: string;
    file_path: string;
    patch: string;
    risk_score: number;
  }) {
    return realApplyPatch(args);
  },
});

TAME checks the tool call before execution. If a policy blocks the call, the SDK throws TameToolError and the protected tool is not executed.

Lifecycle and model telemetry

await tame.recordAgentEvent({
  event_type: "model.completed",
  trace_id: "trace_123",
  payload: {
    provider: "your-provider",
    model: "your-model-name",
    outcome: "success",
  },
});

Supported event types are agent.started, agent.completed, agent.failed, model.requested, model.completed, and model.failed. Send metadata only: never prompts, completions, credentials, or raw customer data.

Notes

• Send only sanitized tool arguments and evidence.
• Do not send secrets, tokens, private keys, or full customer records.
• Keep policy decisions deterministic. AI explanations happen after incidents are stored.