I wanted to share a VS Code Extension I made for those who use both Codex + Claude - Hydra
I wanted to share a thing I have been building. It is a free, open source VS Code extension that puts Claude Code and Codex in one shared room so they plan, build, and review the same task together, with you approving the parts that matter.
I use both CLIs every day and got tired of being the copy paste bridge between them. Ask one, paste into the other to check it, move diffs around, etc.. Now they talk to each other directly, plans for further open source models are in the works..
Basically how it works..
You send one message to the room and both agents get it.
They discuss briefly and try to agree on a plan. One opens, the other reacts.
One agent builds. You choose which, or let it decide.
The other reviews the diff and calls out anything that looks wrong.
You verify, then it loops back for another round if needed.
Auto mode is the part of this I really like the most now. Every agent turn ends with a small structured decision packet: what it did, what it recommends next, any blockers, and a safe default action. With autopilot on, Hydra takes that safe default and advances the loop on its own, discussion into build, build into verify, verify into review, and around again. It runs your test or build command automatically after each build. The thing that makes it usable is the risk gate. The moment a step looks dangerous, like a push, a delete, a deploy, it stops and hands the decision back to you instead of running it. So you can let it grind through the boring middle of a task and only jump in for the calls that actually need you.
Other stuff it does..
Per phase model and reasoning selection. Run a cheap model for the discussion and a strong one for the build. 0.6 adds the current models: Fable, Sonnet 5, GPT 5.6, Opus 4.x, and Haiku.
A live cost meter, plus an optional hard spend cap per session, read from a local usage log.
Parallel mode, where both agents build and review at the same time instead of taking turns.
Project memory. Hydra compiles what happened into a small wiki based on Andrej Karpathy's wiki model, under .hydra and feeds the relevant parts back into later prompts, so a long running project keeps its context across turns and sessions.
File attachments. Drop a file into a turn and it gets copied locally with a bounded preview added to the prompt.
Live streaming, so you watch each reply as it is written instead of waiting for the whole block.
An optional visible terminal mode if you would rather see the agents working in a real shell.
Telegram notifications. If a decision needs you while you are away from the desk, it pings your phone and you can answer from there. 0.6 added a per sender allowlist for that inbound path.
Prompt templates, a workspace state cleaner, and a doctor command that checks your setup.
Everything stays on your machine. It drives the real Codex and Claude Code CLIs you already have, using your existing logins, and writes the full transcript, the plan, and every decision into a .hydra folder you can read.
The code is MIT licensed and lives here: https://github.com/Geraldlol/hydra If you want to help, issues and pull requests are open. Bug reports, rough edges, feature ideas, and feedback are all genuinely super useful right now. I have no idea what the fuck I'm doing. I am the only person on it so far, so more hands and more testing would go a long way.
You can install by searching "Hydra Agents" in the VS Code extensions panel, or here: https://marketplace.visualstudio.com/items?itemName=geraldlol.vscode-hydra-room
Requirements: the Codex CLI and Claude Code CLI installed and logged in. I develop on Windows so that is the most tested path.
Happy to answer anything about the orchestration, the decision packets, the review handoffs or anything else.
Thanks for taking a look and reading!