anthropic-gateway

A local gateway that speaks the Anthropic Messages API and forwards every request to any OpenAI-compatible provider (kilo, OpenAI, DeepSeek, Groq, local servers, ...). It forces every request to a single configured upstream model, which makes it work with clients that can only talk to Anthropic — including the Claude Code GUI (desktop app) via its custom gateway (3P / enterprise) mode, Claude Code CLI, or any Anthropic-SDK tool.

Claude Code GUI/CLI  ──(Anthropic format)──▶  anthropic-gateway (127.0.0.1:3080)
                                                    │  translate + force model
                                                    ▼
                                        OpenAI-compatible provider (/v1/chat/completions)

Install

npm install                 # inside this folder, once
npm i -g .                  # optional: real global install (needs an admin shell
                            # if node lives under C:\Program Files)

Without admin rights a C:\Users\<you>\bin\anthropic-gateway.cmd wrapper is used instead — both give you the anthropic-gateway command (alias: ag). Open a new terminal after installing so PATH picks it up.

Quick start

anthropic-gateway setup     # interactive wizard (or flags, see below)
anthropic-gateway start     # run the gateway using the default profile

Config, pid and log live in ~/.anthropic-gateway/ (config.json, gateway.pid, gateway.log).

Then point your client at:

• Base URL: http://127.0.0.1:3080 (or the port you configured)
• API key / x-api-key / Authorization: Bearer: the proxy token printed by setup

Non-interactive setup

anthropic-gateway setup --provider kilo --api-key YOUR_KILO_KEY --model kilo-auto/free --yes
anthropic-gateway setup --base-url https://api.openai.com/v1 --api-key sk-... --model gpt-5 --port 3090 --token mytoken --yes

Using multiple profiles

After setting up multiple profiles (e.g., one for kilo, one for OpenAI), you can:

• List all defined profiles: anthropic-gateway profiles
• Start a specific profile: anthropic-gateway start openai
• Set the default profile for future start commands: anthropic-gateway select kilo
• Then anthropic-gateway start will use the kilo profile without specifying it.

Commands

| Command | Description | | --- | --- | | anthropic-gateway setup | Create/overwrite ~/.anthropic-gateway/config.json with a profiles array; tests connectivity (--full-logging to enable full-body logging) | | anthropic-gateway start [profile] | Start the gateway using the given profile name; if omitted, uses defaultProfile from config | | anthropic-gateway stop | Stop the gateway | | anthropic-gateway status | Pretty status: version, pid, upstream, uptime, request/token totals, full-logging state | | anthropic-gateway fulllogging on\|off | Toggle full request/response logging (restart to apply) | | anthropic-gateway autostart on\|off | Launch at Windows logon (Startup folder) | | anthropic-gateway select <profile> | Set the default profile for future start commands | | anthropic-gateway profiles | List all defined profiles in the config file | | anthropic-gateway version | Print the version (also --version / -v) |

config.json

See config.example.json. The config now contains a profiles array and a defaultProfile. Each profile object includes: providerName, upstreamBaseUrl, upstreamApiKey, upstreamModel, port, proxyTokens (accepts several), advertisedModels (models shown to Claude Code model discovery — must look Anthropic-ish, e.g. claude-sonnet-4-5 or sonnet), toolFormat, fullLogging, logFile. Only the profile named by defaultProfile (or the first profile if unspecified) is used by anthropic-gateway start unless a profile name is given explicitly.

Logging

gateway.log (in ~/.anthropic-gateway/) records, per request, the tokens sent/received and a running session total (input / output / total since the gateway started).

To also capture the full request and response bodies (useful for debugging upstream mismatches), enable full logging:

anthropic-gateway fulllogging on     # or: setup --full-logging
anthropic-gateway stop && anthropic-gateway start   # restart to apply

Full bodies are written to the log file only (not the console), so streaming stdout stays clean. Turn it off the same way with fulllogging off.

status shows the same token totals plus version, pid, uptime and the full-logging state:

Status                 UP
Version                1.2.0
PID                    23628
Listen                 127.0.0.1:3080
Upstream model         Qwen/Qwen3.8-27B
...
Requests               42
Tokens (in/out/total)  118300 / 9210 / 127510

Using it with Claude Code GUI (desktop app)

1. Keep the gateway running (anthropic-gateway start, or autostart on).
2. In Claude desktop: Settings → custom provider / Claude Code Setup → base URL http://127.0.0.1:<port>, API key = your proxy token.
3. Run its connection test, restart Claude, open Claude Code — the model selector lists the advertisedModels, and every chat is served by your upstream model.