Goose

Goose is an open-source, on-machine AI agent — desktop app, CLI, and API — governed by the Agentic AI Foundation (AAIF). It talks to any OpenAI-compatible endpoint, so you can add Router9 as a custom provider without any code changes.

Everything below was verified against the Goose CLI (goose 1.45). Run goose --version to check yours; older builds may label the menus slightly differently.

Model configuration

Goose reaches an OpenAI-compatible endpoint in either of two ways. Pick one.

Option A — built-in openai provider (quickest, cross-platform)

Goose's bundled openai provider already speaks Router9's format; you only override the host. Set these environment variables and Goose sends every request to Router9.

macOS / Linux:

export GOOSE_PROVIDER=openai
export OPENAI_HOST=https://api.router9.com      # host only — Goose appends /v1/chat/completions
export OPENAI_API_KEY=sk-r9k-your-key-here
export GOOSE_MODEL=deepseek/deepseek-v4-flash   # any id from GET /v1/models

Windows (PowerShell) — setx persists the values; open a new terminal afterwards:

setx GOOSE_PROVIDER openai
setx OPENAI_HOST https://api.router9.com
setx OPENAI_API_KEY sk-r9k-your-key-here
setx GOOSE_MODEL deepseek/deepseek-v4-flash

OPENAI_HOST is the host only — Goose appends /v1/chat/completions itself.

Option B — named custom provider (router9)

If you keep Goose pointed at other providers too, register Router9 by name. Goose reads custom providers from JSON files under its config directory, in a custom_providers/ subfolder:

Run goose info to print the exact config directory on your machine. Create router9.json there:

{
  "name": "router9",
  "engine": "openai",
  "display_name": "Router9",
  "description": "Router9 flat-rate LLM gateway",
  "api_key_env": "ROUTER9_API_KEY",
  "base_url": "https://api.router9.com/v1/chat/completions",
  "models": [
    { "name": "deepseek/deepseek-v4-flash", "context_limit": 200000 }
  ],
  "supports_streaming": true,
  "requires_auth": true
}

Note base_url here is the full chat-completions URL (unlike OPENAI_HOST, which is host-only). Provide the key through the variable named in api_key_env:

export ROUTER9_API_KEY=sk-r9k-your-key-here

You can also create the same entry interactively: run goose configure, choose Configure Providers → Add custom provider with compatible API.

Running Goose

Non-interactive (one-shot), naming the provider and model explicitly:

goose run --provider router9 --model deepseek/deepseek-v4-flash -t "Summarize README.md"

Interactive session (uses your configured/default provider and GOOSE_MODEL):

goose session

Earlier docs showed goose session start --provider …. That subcommand does not exist — goose session has no start verb and no --provider flag. Select the provider with goose configure, the GOOSE_PROVIDER env var, or goose run --provider.

Choosing models

GOOSE_MODEL / --model take the model id Router9 routes on. Use the exact ids from GET /v1/models — Router9 ids are provider-prefixed slugs, so short names like gpt-4o return 404 model_not_found:

curl https://api.router9.com/v1/models \
  -H "Authorization: Bearer sk-r9k-your-key-here"

The exact set depends on your deployment — see Models.

If your administrator has configured an auto route, you can set the model to auto and let Router9 pick. auto is not a built-in alias — when no such route exists it returns 404 model_not_found, so check GET /v1/models first.

MCP Setup

Goose is built around extensions (MCP servers). Add Router9's hosted MCP server as a Remote Extension (Streamable HTTP) to expose Router9 Skills as tools.

Interactively: goose configure → Add Extension → Remote Extension (Streamable HTTP), then answer:

Or add it by hand to config.yaml (same directory as above, goose info prints it):

extensions:
  router9:
    enabled: true
    type: streamable_http
    name: router9
    uri: https://api.router9.com/v1/mcp
    headers:
      Authorization: "Bearer sk-r9k-your-key-here"
    timeout: 300
    bundled: null
    description: Router9 hosted skills

Start a session and the Router9 tools (image_generate, memory_list, …) are available. See Install MCP Server for the full tool list.

Skills (optional): Router9 also offers image and audio Skills as agent tools. MCP is the recommended way to expose them where your harness supports it; otherwise call them directly as REST endpoints (/v1/image/*, /v1/audio/*, /v1/search/) with your sk-r9k- key — see the API Reference.

Tips

• Router9's flat monthly pricing suits Goose's long, autonomous sessions — no per-token surprises.
• Keep your key in an environment variable (OPENAI_API_KEY / ROUTER9_API_KEY) rather than committing it to a provider JSON or config.yaml.
• goose info prints the resolved config paths per OS; goose configure edits them for you.