Client Setup¶

Point your coding tools at Janus. Most OpenAI/Anthropic clients use http://localhost:20128/v1 (adjust the port if you changed it). Gemini-native and Ollama-native clients use the host root http://localhost:20128.

Tool Setup page

The dashboard at /dashboard/tools shows copy-paste environment variable cards for Claude Code, Codex, Cursor, and Cline — tailored to your server URL and auth settings. For Codex CLI, prefer the config.toml recipe below so requests hit /v1/responses.

Claude Code¶

export ANTHROPIC_BASE_URL=http://localhost:20128/v1

If require_api_key is enabled in your config (or toggled from dashboard Settings):

export ANTHROPIC_API_KEY=sk-janus-yourkey

Claude Code sends Anthropic-format requests. Janus translates and routes them to any configured provider.

Codex CLI (Responses API)¶

Codex CLI speaks the OpenAI Responses API (POST /v1/responses), not Chat Completions. Janus exposes that endpoint natively.

Add a provider in ~/.codex/config.toml:

model = "openai/gpt-4o"
model_provider = "janus"

[model_providers.janus]
name = "Janus"
base_url = "http://localhost:20128/v1"
env_key = "JANUS_API_KEY"
wire_api = "responses"

export JANUS_API_KEY=sk-janus-yourkey   # if require_api_key is on

base_url must include /v1 — Codex appends /responses. Use any Janus prefix/model or combo name as model. Streaming uses named SSE events (response.created … response.completed); there is no [DONE] sentinel.

Older guides that only set OPENAI_BASE_URL may still work on some Codex builds, but config.toml + wire_api = "responses" is the reliable path.

Cursor and other Chat Completions tools¶

export OPENAI_BASE_URL=http://localhost:20128/v1

If require_api_key is enabled:

export OPENAI_API_KEY=sk-janus-yourkey

Cursor, Cline (OpenAI Compatible mode), and similar tools send OpenAI Chat Completions to POST /v1/chat/completions. Janus translates and routes.

In Cursor: set the OpenAI base URL / custom endpoint to http://localhost:20128/v1 and the API key to your Janus key. Pick models as prefix/model (or a combo name) in the model picker when the tool allows custom IDs.

Cline (VS Code)¶

In Cline settings:

Set API Provider to "OpenAI Compatible"
Set Base URL to http://localhost:20128/v1
Set API Key to your Janus key (or any value if require_api_key is off)

Generic OpenAI-Compatible Clients¶

Any tool that accepts a custom OpenAI base URL works with Janus:

Base URL: http://localhost:20128/v1
API Key: Your Janus key (if auth enabled)
Model: Use prefix/model format (e.g., openai/gpt-4o, anthropic/claude-sonnet-4-20250514) or a combo name

Creating an API Key¶

janus keys create --name "my-key"
# API Key (save this — shown once): sk-janus-a1b2c3d4...
# ID: 1  Name: my-key
# Login: yes
# Models: all

Optional scopes:

# API-only key (cannot open the dashboard), limited models, $5/day budget
janus keys create --name "agent" --no-login --models "openai/*,my-combo" --daily-budget 5
janus keys update 1 --login --clear-models   # restore full access later

The full key is shown once. Use it in the Authorization: Bearer <key> header or as your OPENAI_API_KEY / ANTHROPIC_API_KEY. Empty model allowlist means all models; patterns may be exact IDs or prefix/*.

Gemini CLI and Gemini-native tools¶

Point Gemini SDKs, Gemini CLI, and other Gemini-native clients at Janus. The base URL is the host root (http://localhost:20128) — the Gemini surface is mounted at /v1beta/..., not under /v1:

export GOOGLE_GEMINI_BASE_URL=http://localhost:20128
export GEMINI_API_KEY=sk-janus-yourkey

Some clients use GOOGLE_API_KEY or x-goog-api-key instead of GEMINI_API_KEY; Janus accepts the standard Gemini auth styles: x-goog-api-key header or the ?key= query parameter (when require_api_key is on).

Tools send requests to POST /v1beta/models/{model}:generateContent (and the streaming streamGenerateContent variant). Use the prefix/model convention in the model path (e.g. openai/gpt-4o or gemini/gemini-2.5-pro) so Janus can route it — the upstream provider does not need to be Gemini.

Cursor typically uses the OpenAI Chat Completions path above, not this Gemini surface. Use Gemini CLI / Google GenAI SDKs here.

Ollama-Only Tools¶

Point Ollama-native clients at Janus (base URL http://localhost:20128, not under /v1):

export OLLAMA_HOST=http://localhost:20128

If require_api_key is enabled:

export OLLAMA_API_KEY=sk-janus-yourkey

Clients use POST /api/chat (messages), POST /api/generate (bare prompt), or POST /api/show (model metadata handshake). GET /api/tags lists routable models (filtered by the key's model allowlist when scopes are set); GET /api/version is available for client handshakes. Streaming defaults to on with application/x-ndjson output.

Model Naming¶

Models are referenced as {prefix}/{model}:

You send	Janus routes to
`openai/gpt-4o`	OpenAI provider, model `gpt-4o`
`anthropic/claude-sonnet-4-20250514`	Anthropic provider, model `claude-sonnet-4-20250514`
`gemini/gemini-2.5-pro`	Gemini provider, model `gemini-2.5-pro`
`copilot/gpt-4o`	GitHub Copilot OAuth provider
`codex/o3`	ChatGPT Codex Responses API (when configured)
`claude/claude-sonnet-4-20250514`	Claude Code OAuth account (when configured)
`best-effort`	Combo — tries each model in the combo chain

See Combos for fallback chain configuration.

Subscription / OAuth providers¶

GitHub Copilot supports Connect GitHub Account (device flow) in the Providers UI. For Codex, Kiro, Antigravity/Gemini CLI, and Claude Code OAuth, paste either:

a bare access token into the API key field, or
a JSON credential blob: {"access_token":"...","refresh_token":"...","expires_at":1710000000}

When refresh_token is present Janus refreshes access tokens automatically before expiry. Claude Code clients also get tool-dedupe + wire-shape normalization when talking to Anthropic-format upstreams.

Remote / Docker setups¶

When Janus runs on another machine or in Docker with host: 0.0.0.0, replace localhost with the host's address. Enable require_api_key and use a dashboard-created key. The dashboard requires login from non-loopback clients — see Dashboard — Authentication.