Rysh vs Codex CLI: a coding agent in your repo vs an agent platform in your company
Codex-class agents nail the repo loop. The question nobody asks: what happens to agent work that isn't a pull request?
Here's a comparison that's mostly not a comparison.
Codex CLI is OpenAI's coding agent for the terminal — you point it at a repo, it plans, edits, runs tests, and iterates until the task is done. By all accounts it's very good at exactly that.
Rysh is not trying to out-code it. Rysh is an agentic dev platform — the workspace layer where many agents (and the humans working with them) live, coordinate, get observed, and get governed. Different layer of the stack. But people evaluating "AI in the terminal" put them side by side, so let's do it honestly.
What Codex-class agents get right
The repo loop: read the codebase, propose a change, run the tests, fix, repeat. One developer, one repository, one conversation. If your definition of "AI at work" is "code gets written faster," a focused coding agent is the shortest path, and OpenAI's iteration speed on Codex is real.
If that's your whole problem — stop reading, install one, be happy.
The ceiling: agent work that isn't a pull request
Now zoom out from the repo to the company. Ask what else needs doing this week:
- ops runbooks and incident response
- a browser task someone repeats every morning
- customer questions arriving on email and website chat
- an internal API someone wants agents to call safely
A coding agent has no opinion about any of this. It lives in the repo. Every one of those tasks either stays manual or grows its own separate AI tool — and now you have four vendors, four permission models, and zero shared visibility.
Rysh's answer is structural: one engine, every surface. The same markdown skill file that defines a terminal agent runs as a Slack/email/website-chat "humanoid," drives a browser agent, and calls internal APIs through Forge-generated tools — with the same governance everywhere: every tool call visible in a pane, dangerous operations gated behind human approval, secrets tokenized client-side before they touch the provider's wire (##snat).
The honest table
| Codex CLI | rysh | |
|---|---|---|
| Core job | Autonomous coding in a repo | Workspace + runtime for many agents across surfaces |
| Model | OpenAI models | Claude — bring your own Anthropic key |
| Scope | The repo loop | Terminal, agents, browser, Slack/email/webchat, internal APIs |
| Multi-agent | Its own internal orchestration | Named agents in coordinated panes; @agent, ##pane listen |
| Collaboration | Single user | Live pane sharing, view or control |
| Governance | Its own sandbox/approval model | Approval gates + client-side secret tokenization, uniform across surfaces |
| Runs on your infra | CLI on your machine | Self-hostable end to end, your keys |
One row in that table deserves extra honesty: the model. Rysh runs on Claude. Not a hedge, not multi-model marketing — the provider layer is pluggable, but Claude is the live backend, and you bring your own key. If your org is standardized on OpenAI models, that's a genuine reason Codex fits you better today. I'd rather tell you that here than have you find out in a demo.
"Can't we just run several Codex sessions in tmux?"
You can, and plenty of people do. What you get is N independent conversations that can't see each other. What you don't get: agents that read each other's output (##pane listen), a session you can hand to a teammate live (##share pane control), headless remote control (rysh send), a shared todo list and context store across agents, or one audit surface for everything the agents did. The multiplexer part of rysh is table stakes; the coordination part is the product.
When to pick Codex CLI
- You're a solo dev or a small team, and the bottleneck is strictly "code gets written."
- Your company runs on OpenAI models and that's not changing.
- You want a tool with zero workspace concepts to learn — one repo, one agent, go.
All legitimate. A focused tool that does one loop brilliantly beats a platform you don't need yet.
When to pick rysh
- Agent work at your company has outgrown the repo: ops, browser tasks, customer channels.
- More than one agent (or more than one human) needs to work in the same space, visibly.
- Security asks "what exactly can this thing do, and who approved it?" — and you want one answer, not four.
- You want the whole stack on your own infra with your own key.
And if the real answer is "both" — rysh panes are real terminals. A Codex-class CLI runs inside one just fine, the same way Claude Code does, and picks up live sharing and session persistence for free.
The category question underneath all of this: do you want AI that makes a developer faster, or a platform that makes the company AI-native? The first is a tool choice. The second is an architecture choice — and it's the one rysh exists for.
Rysh is in private beta. We're recruiting design partners — free access, a direct line to the founder, roadmap influence → apply here.
More in this series: Rysh vs Claude Code · Your company doesn't need another AI subscription