Gemini × OpenCode

Explore Gemini Web tool calling from OpenCode.

A pinned Go adapter, a bounded request queue, and reproducible tests for developers investigating the boundary between web chat and local coding tools.

Quick start · Setup guide · Architecture · Results · Troubleshooting

---

What is this?

Gemini × OpenCode connects OpenCode CLI/Desktop to a patched Gemini Web2API Go backend through a local 9router gateway and request queue. The bridge converts a validated JSON action plan into tool calls that OpenCode can execute.

This repository collects the adapter source, patch recipe, Windows launcher references, and tests from the J.A.R.V.I.S Agents pilot. It is intended for developers studying tool emulation, account binding, streaming, and context limits.

Project status: experimental source archive. Runtime development was cancelled on October 26, 2026. Publishing this repository does not restart that work. This is a Gemini Web adapter, not Google's official Gemini CLI. Pilot model names are routing aliases, not a promise of current Google availability.

Why explore it?

Area	What you can inspect
Tool-call bridge	JSON plan validation, tool selection, argument checks, tool-result correlation, and final-answer handling
Account binding	Identity checks carried through tokens, headers, RPCs, uploads, conversations, and media
Request queue	Bounded concurrency, cancellation, deadlines, and explicit refusal handling
Context preflight	A UTF-8 byte budget checked before uploads or streaming begin
Model identity	A guard against explicitly reported model downgrades before output is emitted
Reproducible patches	An exact upstream revision and 62 source-file hashes checked before any patch is written

How it fits together

flowchart LR
  O["OpenCode CLI / Desktop"] --> R["9router<br/>127.0.0.1:20128"]
  R --> Q["Request queue<br/>127.0.0.1:20832"]
  Q --> G["Patched Gemini Web2API Go<br/>127.0.0.1:20831"]
  G --> W["Gemini Web"]

Loading

The Go bridge proposes structured actions. OpenCode owns tool execution and permissions. The queue forwards requests; it never executes model-generated commands.

Quick start

Start with the local test suite. It requires Node.js 22+, no npm dependencies, and no Google credentials.

git clone https://github.com/Sonness1304/gemini-opencode.git
cd gemini-opencode
npm test

On Windows with PowerShell 7 and the .NET Framework C# compiler:

npm run test:windows

To rebuild the adapter and connect OpenCode, follow the setup guide. The Go files here are patches for a pinned upstream application, not a standalone Go module. The repository does not include an installer or configured gateway.

What has been verified?

Check	Result	Scope
Queue, protocol, preflight, and repository integrity tests	14 passed	Local fixtures; no live model calls
Windows launcher tests	2 passed	Mocked startup/browser callbacks
Patched Go test suite	100 test/subtest passes	One live upload probe skipped; see verification record
Patched Go build	Passed	Go 1.26.8 on Windows
Historical single-tool round trip	Observed passing runs	Tool call → newly generated nonce → exact final answer
High-only multi-tool chains	Not accepted	Mixed model reports, premature finals, and malformed replies remain

These are separate checks, not a combined benchmark. The evidence guide includes failures and explains what each result proves.

Know the limits

• Tool calling is prompt-based emulation, not Google's native function-calling API.
• HTTP 200 or an OpenCode succeeded status does not prove that a requested file was written or a tool chain completed.
• A historical 500000 context setting is client metadata. The pilot's 128000 prompt cap counts UTF-8 bytes of the serialized prompt, not tokens.
• Account login, quota, session expiry, reboot recovery, and media support require their own live verification.
• The repository contains no credentials, account database, browser session, chat history, or prebuilt executable.

Documentation

Guide	Start here when you want to…
Setup	Rebuild the pinned Go application and configure the local chain
Architecture	Understand the adapter, queue, and execution boundaries
Troubleshooting and FAQ	Investigate context errors, model mismatches, or incomplete tool chains
Evidence	Review historical results and unresolved limitations
Publication checks	Inspect the exact checks performed for this export
Contributing	Propose a focused improvement with reproducible evidence
Source and licensing	Trace upstream revisions, original files, and local edits

Explore or contribute

Run the tests, read the bridge, and compare its contracts with your own client. Useful contributions include small reproductions of protocol failures, clearer setup instructions, and tests that catch false success reports. Read CONTRIBUTING.md before opening a pull request.

License

MIT. Upstream copyright and license notices are preserved in THIRD_PARTY_NOTICES.md.