-
Notifications
You must be signed in to change notification settings - Fork 0
FAQ en
Languages: EN · 中文 · 日本語 · Deutsch · Tiếng Việt · 繁體中文 · Home
High-frequency questions for AIGateway users. Also see Troubleshooting and Getting Started.
A local AI model manager + traffic gateway. It connects multi-provider LLM APIs with tools such as ChatGPT/Codex, Claude Code, OpenClaw, and Harness:
- Save tokens — usage by provider / model
- Cheaper vendors — manage many APIs in one place
- One setup, many tools — tools only point at the local gateway
- Hot routing & resilience — virtual model groups + priority failover (e.g. 429 / quota)
V2 is take over once, route forever: tools’ base_url points at the gateway once; switch models/providers inside AIGateway—less file editing, restarts, and env conflicts.
AIGateway runs locally. Requests go to the upstream APIs you configure. Config and usage default to ~/.codex-manager/. Protect your API keys.
macOS (Apple Silicon / Intel), Windows, Linux. Packages: Releases.
Desktop releases: no. Building from source needs Go and a frontend toolchain.
http://127.0.0.1:18080/v1
Check status under Gateway.
Prefer Apps → one-click takeover. Manual:
- Base URL:
http://127.0.0.1:18080/v1 - API key: placeholder
aigatewayis fine (local gateway) -
model: model group name (same as in Models)
Yes. Traffic hits the local proxy first; quitting AIGateway stops the gateway.
Providers → preset (DeepSeek, SiliconFlow, Ollama, OpenAI, Qwen, …) → paste key → fetch/enable models. Most cloud vendors only need a key.
| Mode | Meaning | When |
|---|---|---|
| Standard OpenAI | Normalize request/response | Compatible cloud APIs, unified shape |
| Passthrough | Forward body with minimal rewrite | Special protocols / debugging |
Prefer Standard OpenAI first; try Passthrough if something breaks.
Control whether models participate in routing. Disabled models are not failover candidates.
Group same/equivalent models across providers. Clients use one model name; the gateway picks a channel by priority.
Upstream errors such as 429, quota, account issues (e.g. 401). If all fail → model_group_all_exhausted (OpenAI-style JSON).
Limited: first-byte / HTTP errors can switch; after streaming starts, seamless mid-stream switch is not available. See Troubleshooting.
In Models, drag providers in a group—higher = preferred.
ChatGPT (Codex), Claude Code, OpenClaw, Harness. Other OpenAI-compatible clients: set base URL manually.
Locate tool config → backup → point base_url (and needed provider fields) at the local gateway.
Uninstall / restore on the app card, or use ~/.codex-manager/backups/.
Prefer models.providers.aigateway. Wrong root-only fields may skip the gateway. One-click writes the recommended structure.
See Troubleshooting: rename reserved openai/ollama; use inline api_key = "aigateway".
Usage tab; stored in SQLite (aigateway.db). Paths: Data Paths.
No. Keys stay local. Redact keys when filing issues.
Popup languages: Simplified Chinese, Traditional Chinese, English, Japanese, Korean, German, Vietnamese, Thai.
- Providers — channels & keys
- Models — virtual groups & failover
- Apps — tool takeover
- Gateway — status
- Usage — tokens
- First v2 start: JSON → SQLite migration
- Re-check gateway + takeover status
- Codex ID / api_key rules as above
Known limits: mid-stream failover; some UI fields still evolving—see docs/PRD_V2_CHECKLIST.md and release notes.
- Issues: https://github.com/vincent1986/AIGateway/issues/new/choose
- License: MIT © Mars Waller
| Item | Value |
|---|---|
| Default base URL | http://127.0.0.1:18080/v1 |
| Local API key placeholder | aigateway |
| Data dir | ~/.codex-manager/ |
| Main DB | aigateway.db |
| All-channels-exhausted code | model_group_all_exhausted |
AIGateway · MIT License · Issues · Wiki Home · EN · 中文 · 日本語 · DE · VI · 繁中
AIGateway Wiki
Language / 语言
English
中文(简体)
日本語
Deutsch
Tiếng Việt
繁體中文
Links