pi-9router-ext

Pi Coding Agent extension for 9router — an open-source AI routing proxy.

Install: pi install npm:pi-9router-ext

Connects Pi to your 9router instance via its OpenAI-compatible API, with dynamic model discovery and interactive configuration.

Features

• Auto-discovery — Fetches available models and combos from 9router on startup
• Dynamic provider — Registers 9router as a Pi provider with live model list
• Pi-native streaming — Uses Pi's built-in OpenAI completions provider without overriding other providers
• Status commands — /9router-status, /9router-models, /9router-config, /9router-reasoning, /9router-reload
• Manual reasoning toggle — Optionally expose Pi thinking levels and send reasoning_effort to 9router
• Web tools — Exposes 9router web search/fetch routes as LLM-callable Pi tools
• User-wide persistence — Configuration survives new Pi instances via ~/.pi/agent/9router-config.json
• Routing detection — Captures upstream model info from response headers when available

Installation

npm (Recommended)

pi install npm:pi-9router-ext

Via local path

# Clone or download this repo
git clone https://github.com/irfansofyana/pi-9router-ext.git

# Install locally
pi install /path/to/pi-9router-ext

# Or try without installing
pi -e /path/to/pi-9router-ext

Manual / development install

Prefer installing the package directory instead of copying individual source files. The extension may include multiple files, tools, commands, and helper modules; installing the directory lets Pi read the package manifest and load everything declared there.

# Clone or update the repo
git clone https://github.com/irfansofyana/pi-9router-ext.git
cd pi-9router-ext

# Install this working tree into Pi
pi install "$PWD"

# Reload Pi, then use /9router-config

Avoid copying only src/index.ts into ~/.pi/agent/extensions. That bypasses package.json metadata and can miss companion files added by future features.

Configuration

Environment Variables

| Variable | Default | Description | |----------|---------|-------------| | NINE_ROUTER_BASE_URL | http://localhost:20128 | Your 9router instance URL | | NINE_ROUTER_API_KEY | — | API key if 9router has REQUIRE_API_KEY=true | | NINE_ROUTER_ENABLE_REASONING | false | Set to true/1/on to expose Pi thinking levels for 9router models |

Set them in your shell profile or prefix your pi command:

NINE_ROUTER_BASE_URL=http://my-vps:20128 NINE_ROUTER_API_KEY=nr-... pi

Interactive Configuration

Use the /9router-config command inside Pi to open a configuration menu for connection settings, reasoning, web defaults, and status/routes. This is saved to ~/.pi/agent/9router-config.json and is shared by new Pi instances. Environment variables still take precedence when set.

/9router-config

Use /9router-reasoning to quickly enable or disable reasoning without changing the base URL or API key.

Web search/fetch defaults are configured from discovered GET /v1/models/web routes inside /9router-config. Direct provider routes such as brave/search and tavily/fetch are supported, as are 9router web combos.

Usage

Selecting a 9router Model

After the extension loads, 9router models are available in Pi's model picker under the dedicated 9router/ provider namespace:

/model 9router/cc/claude-opus-4-7

Built-in providers remain separate. To use Ollama Cloud, OpenRouter, opencode-go, etc., select their normal provider/model entries (for example ollama-cloud/...), not a 9router/... model.

Or browse interactively:

/9router-models

Reasoning / Thinking Levels

9router's /v1/models endpoint does not currently expose reliable per-model reasoning capability metadata. For safety, this extension keeps reasoning disabled by default.

If your selected 9router route/model supports reasoning, enable the manual toggle:

/9router-reasoning

When enabled, Pi treats 9router models as reasoning-capable. Use Pi's normal thinking controls such as Shift+Tab, --thinking high, or model suffixes like 9router/cx/gpt-5.3-codex:high. Pi sends OpenAI-style reasoning_effort values to 9router (off → none, low, medium, high, xhigh).

Web Search / Fetch Tools

If 9router exposes web routes from GET /v1/models/web, this extension lets the LLM call them through Pi tools:

• ninerouter_web_search — calls POST /v1/search
• ninerouter_web_fetch — calls POST /v1/web/fetch

Tools are always registered. If no matching web route is configured or discovered, they fail with an actionable message telling you to configure web defaults in /9router-config.

Search supports all generic fields exposed by 9router:

{
  "query": "latest pi coding agent extension docs",
  "route": "brave/search",
  "max_results": 5,
  "search_type": "web",
  "country": "US",
  "language": "en",
  "time_range": "week",
  "offset": 0,
  "domain_filter": ["github.com"],
  "content_options": {},
  "provider_options": {}
}

Fetch supports:

{
  "url": "https://example.com",
  "route": "tavily/fetch",
  "format": "markdown",
  "max_characters": 12000
}

route is optional. If omitted, the configured default route is used; if that route disappears, the extension falls back to the first discovered compatible route and records that in tool details/status. Per-call route overrides can be direct routes (brave/search, tavily/fetch) or combo names.

Available Commands

| Command | Description | |---------|-------------| | /9router-status | Show connection status, model count, web route count, and config | | /9router-models | Browse and select from available 9router models | | /9router-config | Menu for connection, reasoning, web defaults, and status/routes | | /9router-reasoning | Enable or disable Pi thinking levels for 9router models | | /9router-reload | Refresh model list and web routes from 9router |

Available Tools

The LLM can call:

• ninerouter_status — check connection status, model list, and web route defaults
• ninerouter_web_search — search the web through 9router
• ninerouter_web_fetch — fetch/extract URL content through 9router

How It Works

┌─────────┐     OpenAI-compatible      ┌──────────┐     ┌─────────────┐
│   Pi    │ ──────────────────────────▶│ 9router  │ ──▶ │  Providers  │
│         │     /v1/chat/completions   │ (proxy)  │     │ (40+)       │
└─────────┘                            └──────────┘     └─────────────┘
     │
     │ 1. Fetches /v1/models and /v1/models/web on startup
     │ 2. Registers as provider "9router"
     │ 3. Uses pi-ai's normal OpenAI implementation for only 9router models
     │ 4. Registers web search/fetch as Pi tools
     │ 5. Leaves built-in/non-9router model switching untouched

Troubleshooting

"Failed to discover models from 9router"

• Check that 9router is running: curl http://localhost:20128/v1/models
• Verify NINE_ROUTER_BASE_URL points to the correct host/port
• If REQUIRE_API_KEY=true, set NINE_ROUTER_API_KEY

"No 9router models discovered"

• 9router may have no active providers. Open the dashboard and connect a provider.
• Use /9router-reload to retry discovery.

"No 9router web search/fetch route is configured or discovered"

• Open the 9router dashboard and connect a web provider such as Brave, Tavily, Exa, or a web combo.
• Run /9router-reload or open /9router-config → Web defaults to refresh /v1/models/web.
• Pick separate default routes for search and fetch if more than one route exists.

Models not showing in /model selector

• Ensure the extension loaded: check for "9router connected" notification on startup
• Run /9router-reload to retry discovery
• Run /reload to refresh extensions

Built-in provider returns 401 after installing this extension

• Make sure the active model is the built-in provider (ollama-cloud/..., openrouter/..., etc.), not a 9router/... route.
• Re-run /login for the built-in provider if its subscription token expired.
• This extension only registers provider id 9router; it does not override built-in providers. If 9router is unreachable, the extension unregisters the 9router provider instead of leaving an empty/broken provider around.

Similar Projects

• omniroute-pi-extension — Pi extension for OmniRoute (a fork of 9router with additional features)

License

MIT