@openagentry/adapter-llm-openrouter

Third concrete LlmAdapter for OpenAgentry — runtime LLM access via OpenRouter (plain fetch against their OpenAI-compatible REST API). Joins @openagentry/adapter-llm-cli (operator/CLI) and @openagentry/adapter-llm-aigateway (runtime/SDK).

Install

npm install @openagentry/adapter-llm-openrouter
# or
pnpm add @openagentry/adapter-llm-openrouter

Quick start

import llm from '@openagentry/adapter-llm-openrouter';

// Set OA_OPENROUTER_API_KEY in env, then:
const models = await llm.listModels();
const out = await llm.generate('Hello!', { model: 'openai/gpt-5.2' });
console.log(out.text);

Configuration

| Env var | Required | Purpose | |---|---|---| | OA_OPENROUTER_API_KEY | yes | OpenRouter API key (https://openrouter.ai/keys) | | OA_OPENROUTER_DEFAULT_MODEL | no | Default model when opts.model unset | | OA_OPENROUTER_BASE_URL | no | Override base URL (default: https://openrouter.ai/api/v1) | | OA_OPENROUTER_APP_URL | no | Sent as HTTP-Referer (opt-in attribution) | | OA_OPENROUTER_APP_TITLE | no | Sent as X-OpenRouter-Title (opt-in attribution) |

Operator vs runtime

This adapter is the runtime flavor — for deployed apps that call the LLM at request time. For build-time/operator-side calls (no API key needed), use @openagentry/adapter-llm-cli.

What's NOT in v0

• Tool/function-calling support
• Vision/image inputs
• Embeddings endpoint
• Provider-routing preferences (route, provider: {...} body params; pass model: 'openrouter/auto' for auto-routing instead)
• Retry logic on transient 5xx
• Cost tracking from response headers
• Cache-Control parsing helpers (no equivalent need on this adapter)

Documentation

• LLM-targeted reference: LLM.md
• License: Apache-2.0
• Issues: https://github.com/timncox/openagentry/issues

Architecture notes

Plain fetch everywhere — no ai / openai / @openrouter/sdk dependency. SSE streaming uses a hand-rolled line-buffered parser (src/sse.ts). Mocking in tests uses pure MSW (no vi.mock shenanigans, since globalThis.fetch is the only external call surface).

The deliberately-different transport shape (CLI shell-out / SDK wrapper / raw HTTP) across the three LLM adapters strengthens the abstraction proof for the category — the same LlmAdapter interface fits all three.