@tepa/provider-anthropic

Anthropic Claude LLM provider for the Tepa agent pipeline.

Install

npm install @tepa/provider-anthropic

Setup

Set the ANTHROPIC_API_KEY environment variable. You can either export it directly:

export ANTHROPIC_API_KEY=sk-ant-...

Or use a .env file with dotenv:

# .env
ANTHROPIC_API_KEY=sk-ant-...

import "dotenv/config";

Usage

import { Tepa } from "@tepa/core";
import { AnthropicProvider } from "@tepa/provider-anthropic";

const tepa = new Tepa({
  tools: [
    /* ... */
  ],
  provider: new AnthropicProvider(),
});

Provider Options

const provider = new AnthropicProvider({
  apiKey: "sk-ant-...", // defaults to ANTHROPIC_API_KEY env var
  maxRetries: 3, // retry attempts on transient failures (default: 3)
  retryBaseDelayMs: 1000, // base delay for exponential backoff (default: 1000)
});

Factory Function

Use createProvider to create providers from a string identifier:

import { createProvider } from "@tepa/provider-anthropic";

const provider = createProvider("anthropic");

Logging

Every LLM call is automatically logged to a JSONL file in .tepa/logs/. You can disable the default file logger, add custom log listeners, or send logs to external services like Prometheus, NewRelic, or Datadog using the onLog() method:

const provider = new AnthropicProvider({ defaultLog: false });

provider.onLog((entry) => {
  externalLogger.send(entry);
});

See @tepa/provider-core for full logging documentation.

Other Providers

Tepa ships with multiple LLM providers — all following the same LLMProvider interface:

• @tepa/provider-gemini — Google Gemini
• @tepa/provider-openai — OpenAI

Implementing Custom Providers

To support a different LLM, implement the LLMProvider interface from @tepa/types:

import type { LLMProvider, LLMMessage, LLMRequestOptions, LLMResponse } from "@tepa/types";

export class MyProvider implements LLMProvider {
  async complete(messages: LLMMessage[], options: LLMRequestOptions): Promise<LLMResponse> {
    // Call your LLM API, passing options.tools if provided
    return {
      text: "response text",
      tokensUsed: { input: 100, output: 50 },
      finishReason: "end_turn",
    };
  }
}

The provider interface is intentionally minimal — one method, clear input/output types.

Native Tool Use

All built-in providers support native tool use. When the executor passes tool schemas via options.tools, the provider forwards them to the LLM's native function-calling API. The LLM returns structured toolUse blocks with pre-parsed parameters — no manual JSON parsing needed. This eliminates escaping errors that occur when LLMs produce tool parameters as free-form text, especially with large content like file writes.