For coding agents

One URL. Full multi-modal capability.

Point your coding agent — Claude Code, Cursor, Cline, Windsurf, OpenClaw — at AIgateway and it can deliver every modality without touching individual provider SDKs. One base URL, one key, every model.

Open /llms.txt →Full catalog (/llms-full.txt)OpenAPI specAgent guide (agents.md)
The three URLs

Give your agent these three
and it configures itself.

No tutorials, no webinars, no 37-tab dashboard. A capable coding agent reads these once and builds.

LLMS.TXT
Human + agent-readable map of every endpoint, modality, and primitive. Start here.
/llms.txt
LLMS-FULL.TXT
Live catalog of every 100+ model with price, context, capabilities — regenerated hourly.
/llms-full.txt
OPENAPI.JSON
Typed OpenAPI 3.1 spec so code-gen tools produce clients without reading prose.
/openapi.json
AGENTS.MD
Agent-specific integration patterns — when to mint a sub-account, when to tag, how to handle errors.
/agents.md
Prompt-paste

Drop this into your coding agent.

We designed this to be pasted verbatim. It's long enough to be unambiguous and short enough to fit in a single system message.

You can call any AI model through AIgateway, a universal AI API.

- Base URL: https://api.aigateway.sh/v1
- Auth: Authorization: Bearer sk-aig-...
- SDK: drop-in for the OpenAI SDK — only change base_url
- OpenAPI spec: https://api.aigateway.sh/openapi.json
- Capability map: https://aigateway.sh/llms.txt
- Model list (live): https://api.aigateway.sh/v1/models

Core endpoints (OpenAI-shaped):
  POST /v1/chat/completions     — text, vision, tool calling, streaming
  POST /v1/embeddings           — embeddings
  POST /v1/images/generations   — image generation
  POST /v1/audio/transcriptions — STT (multipart)
  POST /v1/audio/speech         — TTS
  POST /v1/moderations          — Llama Guard

Aggregator primitives (no other gateway has these):
  POST /v1/sub-accounts — scoped keys + spend caps for per-customer products
  POST /v1/evals        — grade candidate models on a dataset, use eval:<id> as a model alias
  POST /v1/replays      — re-run a past request on a new model, diff cost + latency + output
  GET  /v1/usage/by-tag — per-feature cost attribution via x-aig-tag header

Model IDs use <provider>/<model> slugs. Examples:
  anthropic/claude-opus-4.7, openai/gpt-5.4, google/gemini-3.1-pro,
  moonshot/kimi-k2.6, meta/llama-4-scout-17b-16e-instruct,
  black-forest-labs/flux-2-klein-9b, deepgram/nova-3, baai/bge-m3

Read https://aigateway.sh/agents.md for deeper integration patterns.
Install for agents

Three install lines.
Three host languages.

The OpenAI SDK with our base_url covers chat / embeddings / images / STT / TTS. When the agent needs the aggregator-native primitives — async jobs, sub-accounts, evals, replays, signed URLs, webhook verification — install the typed SDK. The CLI is what you reach for when you want to do all of it from the shell.

PYTHON
aigateway-py
pip install aigateway-py
from aigateway import AIgateway
/integrations/aigateway-python →
NODE
aigateway-js
pnpm add aigateway-js
import { AIgateway } from "aigateway-js"
/integrations/aigateway-node →
CLI
aigateway-cli
npm i -g aigateway-cli
aig init # then: aig call, aig models, aig mcp call
/integrations/cli →
Error messages

Every error tells you
what to do next.

Stripe-grade remediation in the message field. Your agent never has to guess.

StatusTypeWhat to do
401authentication_errorAsk the human for a key at aigateway.sh/signin
402budget_exceededCall PATCH /v1/sub-accounts/:id to raise the cap, or wait for month rollover
404model_not_foundFetch /v1/models and pick the exact <provider>/<model> slug
429rate_limit_errorHonor Retry-After; if persistent, raise the RPM on the key
502provider_errorRetry once; router failover usually catches this
504timeout_errorRetry with a smaller max_tokens