Relay

Chrome extension (MV3) + FastAPI backend. Captures chat context from AI chat tools, cleans it via LLM, stores in Supabase, and injects into other AI chats.

Version: 0.1.0

Supported Sites

• ChatGPT (chatgpt.com, chat.openai.com)
• Claude (claude.ai)
• Perplexity (www.perplexity.ai, perplexity.ai)
• Gemini (gemini.google.com, bard.google.com)
• Copilot (copilot.microsoft.com)
• Grok (grok.com)
• Mistral (chat.mistral.ai)

Quick Start

Prerequisites

• Node.js and npm
• Python 3.12+
• Chrome or Chromium-based browser
• Supabase project
• OpenRouter API key

1. Setup Environment

Create .env from .env.example:

SUPABASE_URL=
SUPABASE_SERVICE_ROLE_KEY=
OPENROUTER_API_KEY=
LLM_API_BASE=https://openrouter.ai/api/v1
LLM_MODEL=openai/gpt-4o-mini
ALLOWED_EXTENSION_ORIGINS=chrome-extension://<your-extension-id>

2. Install Dependencies

python -m venv .venv
.venv\Scripts\python.exe -m pip install -r requirements.txt
npm install

3. Build & Load Extension

npm run build

1. Open chrome://extensions
2. Enable Developer mode
3. Click "Load unpacked"
4. Select the dist/ folder
5. Copy the extension ID
6. Add chrome-extension://<id> to .env ALLOWED_EXTENSION_ORIGINS
7. Restart backend after changing .env

4. Run Backend

npm run backend

Backend starts at http://127.0.0.1:8000. Logs write to uvicorn.log.

Commands

Command	Purpose
npm run build	Compile TypeScript and copy static files to dist/
npm run backend	Start FastAPI backend on 127.0.0.1:8000
npm test	Run Python unittest discover
npm run check:python	py_compile on api/core.py and api/index.py
npm run check	Build + test + python check
npm run generate-icons	Generate PNG icons from SVG (requires sharp)

Architecture

api/
  core.py            # validation, truncation, prompts, parsing, hashing
  index.py           # FastAPI app, routes, OpenRouter, Supabase
  mcp_server.py      # MCP server (FastMCP) at /mcp
extension/
  manifest.json      # MV3 manifest
  popup.html         # popup shell
  popup.css          # popup styles
  injected.css       # content-script injected styles
  assets/            # icons and logo
src/
  background.ts      # service worker message router
  client.ts          # backend fetch client, per-browser client id
  config.ts          # backend URL and supported hosts
  content.ts         # page scraping, save button, import button, context injection
  popup.ts           # popup UI logic
  types.ts           # shared TypeScript interfaces
tests/
  test_core.py       # Python unittest coverage for api/core.py
scripts/
  start-backend.py   # uvicorn launcher
  copy-static.js     # copies extension files into dist/
  generate-icons.js  # generates PNG icons from SVG

Build output goes to dist/. Chrome loads extension files from dist/, not src/ or extension/.

Flow

1. Content script runs on supported AI chat sites
2. Injects a "Save Context" button at bottom-right
3. User clicks button (or uses popup save)
4. Content script scrapes visible chat text
5. Background service worker sends raw chat to backend with X-Relay-Client-Id
6. Backend validates, truncates large chats, sends to LLM for cleanup
7. Backend parses title/content, hashes, dedupes by client + hash
8. Cleaned context stored in Supabase
9. Popup lists saved contexts
10. User selects one → extension injects it into current chat input

Injected format:

You are continuing an existing project.

Here is the full working context:

--- CONTEXT START ---
<saved context>
--- CONTEXT END ---

First, acknowledge you understand this context. Then wait for my next instruction.

Backend API

All context endpoints require:

X-Relay-Client-Id: <client-id>
Content-Type: application/json

GET /

Returns API metadata.

GET /health

Returns config health:

{
  "ok": true,
  "supabase_configured": true,
  "llm_configured": true
}

POST /capture-context

Body:

{ "raw_chat": "..." }

Response:

{
  "id": 1,
  "title": "Project Context",
  "created_at": "2026-05-19T00:00:00Z",
  "deduped": false,
  "truncated": false
}

GET /list-contexts

Returns latest 20 contexts for current client id.

GET /get-context/{context_id}

Returns full context. Missing context returns 404.

MCP Server

Mounted at /mcp on the backend (Streamable HTTP transport). Built with FastMCP.

Tools

All tools require a client_id parameter:

Tool	Description
list_contexts	List up to 20 contexts for a client
get_context	Get full context content by id
capture_context	Save new context from raw chat text (async, uses LLM cleanup)

Client Config

{
  "mcpServers": {
    "relay": {
      "url": "http://localhost:8000/mcp"
    }
  }
}

Test with MCP Inspector

npx -y @modelcontextprotocol/inspector

Connect to http://localhost:8000/mcp.

Environment Variables

Variable	Description
SUPABASE_URL	Supabase project URL
SUPABASE_SERVICE_ROLE_KEY	Supabase service role secret
OPENROUTER_API_KEY	OpenRouter API key
LLM_API_BASE	LLM API base URL (default: https://openrouter.ai/api/v1)
LLM_MODEL	Model to use for cleanup (default: openai/gpt-4o-mini)
ALLOWED_EXTENSION_ORIGINS	Comma-separated extension origins for CORS

Supabase Setup

Run supabase_schema.sql in Supabase SQL editor.

Schema:

create table if not exists public.contexts (
  id bigserial primary key,
  client_id text not null,
  title text not null,
  content text not null,
  content_hash text not null,
  created_at timestamptz not null default now()
);

Indexes:

• contexts_client_hash_idx on (client_id, content_hash) for dedupe
• contexts_client_created_idx on (client_id, created_at desc) for listing

Security: RLS enabled, anon and authenticated privileges revoked, backend uses service role key (bypasses RLS).

Backend Core Rules

Implemented in api/core.py.

Truncation: Raw chats over 60K chars retain first 10K + last 50K with a mid-truncation note.

Cleanup prompt asks model to return:

Title: <short specific title, max 8 words>

Project goal:
...

Constraints:
...

Decisions made:
...

Important information:
...

Open items:
...

Parsing: If output starts with Title:, title extracted (max 120 chars). Fallback: Context - <Mon DD>.

Dedupe: SHA-256 hash of cleaned content, matched by client_id + content_hash.

Security Notes

• Keep SUPABASE_SERVICE_ROLE_KEY only on backend
• Never put service role key in extension code
• Do not commit .env
• Contexts are scoped by generated client id, not user auth
• Client id lives in Chrome local extension storage

Limitations

• No user accounts or auth beyond client id
• No delete or edit/rename endpoint
• No pagination beyond latest 20 contexts
• Scraping depends on third-party site DOM structure
• Backend URL hardcoded in src/config.ts
• Large chats lose middle content after 60K chars
• No linter/formatter configured

License

No license file is currently included.