OmniRoute Pi Extension

A seamless Pi Coding Agent extension that brings OmniRoute — the ultimate AI gateway — directly into your editor environment.

Connect to your local or remote OmniRoute server, browse models, manage combos, check quotas, and intelligently route your Pi queries across 44+ LLM providers.

Features

• 🔮 Wizard-Based Setup: Just run /omni setup inside Pi. No manual JSON editing needed.
• ⚡ Pure HTTP Client: Works securely and seamlessly whether your OmniRoute server is running locally on localhost:20128 or hosted on a remote VPS.
• 🔄 Combo & Model Sync: Instantly push all OmniRoute combos and available models into Pi’s Ctrl+P model picker with full metadata (context windows, max tokens, reasoning support, and vision capabilities).
• 🛠️ Prompt Tool Fallback for Chat-Only Models: Models that do not support native tool_calls can still use Pi tools through prompt-emulated tool calling.
• 🔁 Same /model Workflow: Switch models normally; the extension chooses native tools or prompt tools automatically.
• 📊 Selected Model Status: Status bar shows the currently selected Pi/OmniRoute model ID. For OmniRoute combos, this shows the combo/model selected in Pi, not the underlying provider model that OmniRoute ultimately routed to.
• 🧬 Smart Sorting: Syncing organizes your model list by provider/group (owned_by) for a cleaner Ctrl+P experience.
• 🛠️ Diagnostics & Health: Spot expired tokens, connection failures, or disconnected providers right when Pi starts (management endpoints must be accessible).
• 📉 Quota Management: Live usage tracking mapped directly to OmniRoute's global quota endpoints.

Installation

Install the package directly from NPM:

pi install omniroute-pi-ext-integration

Or install the latest development version directly from GitHub:

pi install git:github.com/md-riaz/omniroute-pi-ext-integration

Getting Started

1. Start Pi:

   pi

2. Run Setup: Once Pi starts, open the command palette and run:

   /omni setup

3. Enter Credentials: Enter your OmniRoute Server URL and API key when prompted. The key is collected before connectivity testing so protected /v1/models endpoints can be verified.
4. Sync Models: Run /omni sync to populate the Ctrl+P list with all your provider models and combos.
5. Switch Models Normally: Use /model as usual. No separate prompt-tools provider is needed.

Prompt Tool Fallback

Some OmniRoute-synced models are chat-only: they can answer text, but they do not return native OpenAI-style message.tool_calls. This is common for web-synced model IDs such as:

cgpt-web/gpt-5.4-pro
chatgpt-web/gpt-5.5
bb-web/gpt-4-turbo
ds-web/deepseek-v4-pro

For these models, the extension keeps the same Pi provider (omni) and /model workflow, but internally switches to prompt-emulated tool calling.

Native tool mode

Used for normal tool-capable models.

Pi agent
  -> omni provider
  -> OmniRoute with native tools: [...]
  -> model returns native tool_calls
  -> Pi executes tools

Prompt tool mode

Used when a model is chat-only or marked as not supporting native tool calls.

Prompt tool mode is intentionally buffered: the extension waits for the full model response before showing text, because it must parse complete <tool_call> blocks before emitting Pi-native tool events.

Pi agent
  -> omni provider
  -> extension renders Pi tools as text instructions
  -> OmniRoute request is sent with tools: []
  -> model writes <tool_call>{...}</tool_call>
  -> extension converts that text back into Pi native toolCall events
  -> Pi executes tools normally

The model is taught this wire format:

<tool_call>
{"name":"read","arguments":{"path":"index.ts"}}
</tool_call>

Tool results are fed back in history as text:

<tool_result tool="read" id="call_123">
...tool output...
</tool_result>

How Prompt Tool Mode Is Detected

Prompt tool mode is enabled when either condition is true:

1. The upstream model metadata contains -web during sync, such as OmniRoute model ID/name, owned_by, or model provider label. This does not change the Pi provider ID, which remains omni.
2. The synced models.json model entry contains:

{
  "tool_calling": false
}

The second check reads raw models.json because Pi's runtime Model object does not preserve custom fields like tool_calling.

Example synced model entry:

{
  "id": "cgpt-web/gpt-5.4-pro",
  "name": "Gpt 5.4 Pro",
  "api": "omni-prompt-tools",
  "tool_calling": false,
  "input": ["text", "image"],
  "contextWindow": 400000,
  "maxTokens": 65535,
  "reasoning": true
}

Model Switching

Use Pi's normal model picker/command:

/model cgpt-web/gpt-5.4-pro
/model codex/gpt-5.2
/model premium

The extension routes automatically:

| Model kind | Detection | Tool mode | |---|---|---| | Web-synced model | Upstream model ID/name, OmniRoute owned_by, or model provider label contains -web | Prompt-emulated tools | | Explicit chat-only model | tool_calling: false in models.json | Prompt-emulated tools | | Normal model | No fallback marker | Native tools |

Commands Reference

| Command | Description | |---|---| | /omni | Dashboard showing server health, active connections, and combos | | /omni sync | Sync your Pi model picker with all healthy OmniRoute instances | | /omni setup | Launch interactive wizard to link Pi with your OmniRoute gateway | | /omni dashboard | Get the direct link to your OmniRoute web interface |

Development

This repo is intentionally small and AI-friendly:

| File | Purpose | |---|---| | AGENTS.md | Required instructions for AI agents editing this repo. | | AI.md | Fast handoff guide for AI agents and future maintainers. | | ARCHITECTURE.md | Data flows and prompt-tool architecture. | | CONTRIBUTING.md | Local setup, test checklist, and contribution rules. | | index.ts | Extension implementation. |

Run TypeScript check:

npm run typecheck

Smoke-test extension import:

npm run smoke

Requirements

• Pi Coding Agent v0.60.0+
• OmniRoute v2.9.0+

License

MIT