Skip to content

GemCoder

GemCoder is a agentic coding tool that lets developers define, run, and improve their own agentic coding harness as code.

Code Video Cerebral Valley

Core trick

Reads per-repo YAML and Markdown files to set agent instructions, skills, and approval policy, then routes each task to cloud or local runtime.

Tags

dev toolsagents

Description

GemCoder is a agentic coding tool that lets developers define, run, and improve their own agentic coding harness as code. Each repository owns its harness through version controlled files. GemCoder orchestrates between Google's Managed Agents API in the cloud and the Antigravity SDK on your laptop, same engine, two runtimes, your choice per task. These files define the agent’s instructions, reusable skills, allowed context, verification commands, approval policy, and runtime routing. GemCoder turns that harness into a repo-aware workflow for the Gemini and Google Managed Agents stack, with safe context mounting, structured task packets, patch-first outputs, local verification, run graphs, and support for local or remote execution. The TUI streams agent steps and tokens live, highlights diffs, and previews patches before applying.
ramizik/gem-coder★ 00 forksPythonCreated May 23, 2026 · last push May 24, 2026

README excerpt

# GemCoder

**GemCoder is the tool for building your *own* coding agent harness** — defined as
code, in your repo, customized to your workspace.

OpenCode is too much. OpenClaw is too little. And the AI vendors push you into a
*closed* harness you can't change. But every project is different — different
language, different rules — so don't settle for someone else's. GemCoder lets you
**build and customize your own**: what the agent does (`AGENTS.md`), the skills it
uses (`.gemcoder/skills/`), what it can touch, how it verifies, and where it runs
(`gemcoder.yaml`) — all version-controlled files you own. **Harness as code.**

And it's not all-or-nothing. **Don't be scared to build your own** — but you don't
have to do it all: run your own models and extend/connect anything (custom tools,
MCP servers), *or* connect to others and reuse existing tools and harnesses. Either
way, you're never locked in.

It turns that harness into a repo-aware developer workflow with project
instructions, skills, structured task packets, patch previews, local
verification, run graphs, and harness evaluation.

**Two backends, one harness.** GemCoder uses the Google agent stack at the right
layer for each task: the **Antigravity SDK** runs the harness **locally** (the
agentic loop on your machine) for lighter tasks, and **ADK 2.0 + Managed Agents**
run it in the **cloud** for bigger ones — two engines, one shared harness
definition. GemCoder's **orchestrator** routes each task (small/quick → local,
bi

Team