pranrichh/gemini-canvas-proxy

Free unlimited Gemini API via Canvas + Chrome extension postMessage bridge. No WebSocket, no Local Network Access issues. OpenAI-compatible endpoint.

Python26 ★Web & Browser AutomationGitHub ↗
Quality: decent 14/24
PAI: watch 0.5
Appraised: 2026-07-24 current

Functional proxy with clear docs and Docker option, but no CI, no dep manifest, and thin adoption.

postMessage bridge bypasses Chrome 142+ Local Network Access blocks that break WebSocket-based Canvas proxies

Canvas auto-injects unlimited, model-scoped API keys when code contains `const apiKey = \"\";` at compile time

Contents

Overview

Verdict

Rating Summary
Quality decent (14/24) Solid documentation and a working Docker path, undercut by Python-only tooling, no manifest, no tests, and thin adoption.
PAI Relevance watch (0.50) A clever browser-bridge pattern worth studying, but the manual extension/Canvas setup and ToS-fragile mechanism keep it out of direct integration.

Quality Assessment

14/24 — maintained / adequately-documented / early-or-minimal

Health: 5/8 (maintained)

Failed:

Passed:

Documentation: 6/8 (adequately-documented)

Failed:

Passed:

Engineering Signals: 3/8 (early-or-minimal)

Failed:

Passed:

PAI Relevance

Dimension Score Assessment
Harvest Value 1 The postMessage-over-WebSocket IPC trick to bypass Chrome's Local Network Access sandboxing is a reusable pattern worth studying, though it maps to no named PAI subsystem directly.
Integration Readiness 1 Exposes a clean OpenAI-compatible HTTP endpoint (localhost:8765) any agent could call, but the Python native-messaging host plus manual Chrome-extension/Canvas paste-in setup isn't a bun add-able CLI.
Overlap Risk 1 Partially overlaps PAI's Interceptor (stealth Chrome automation), which already covers browser-extension-based automation, though this repo's free-inference-via-Canvas trick is a distinct use case.
Gap Fill 1 Touches an area PAI has limited coverage in — sourcing free/alternate LLM inference — though PAI's Agents skill already manages cross-vendor model routing (GPT-5.4, Kimi-K2.6, Claude).

Composite: 0.50

What Next

Landscape Position

Category: Web & Browser Automation

In this category: first entry — no prior appraisals recorded in this category for cross-referencing.

Standing: As the first Web & Browser Automation entry in the vault, it stands alone rather than being compared against ChromeDevTools-MCP or crawl4ai-class tools, which target scraping/DOM control rather than credential-bridging.

Evidence Base

Density: 8/10 — Available: full README (8KB), description, language/license, star/fork/issue counts, creation and last-commit timestamps, release tag. Missing: dependency manifest ("Not available"), no visible CI config, no topics tags, no contributing/limitations doc beyond the inline security note.

Notes

The core mechanism relies on exploiting Google's Canvas auto-injected API credentials — a quota loophole rather than a sanctioned API path. This makes the tool inherently fragile to upstream changes (Google could patch the credential injection at any time) and raises ToS considerations that are out of scope for quality scoring but worth flagging for anyone considering production reliance.