Chat-based coding agents
Vortex vs Claude Code
Claude Code is a terminal coding agent from Anthropic. Vortex is a macOS app plus a headless CLI — both take a job, not a line. Here is where each fits.
Claude Code is a terminal-based coding agent from Anthropic. It works in your project directory and, as of this writing, runs on macOS, Linux, and Windows. You give it a job in the terminal and it works inside the repository.
Vortex is a macOS desktop app plus a headless CLI. It is also job-based, so the unit of work is similar. The differences are the surface you watch it from, the models you can point it at, and where the checkpoint sits before edits land.
Both are legitimate choices. This page is about which one fits which job.
The short version
| Claude Code | Vortex | |
|---|---|---|
| Primary surface | The terminal | macOS desktop app, plus a headless CLI |
| Platform | macOS, Linux, and Windows, as of this writing | macOS 12+; closed beta. Windows and Linux are out of scope for this beta |
| Model choice | Anthropic models, as of this writing | BYO OpenAI, Anthropic, OpenRouter, or any compatible endpoint; optional local GGUF models |
| Plan artifact | As of this writing, not a plan file in the workspace | Plan mode writes .vortex/plans/<name>.plan.md |
| Review checkpoint | You approve actions as the agent requests them in the session | Review mode stages edits as pending patches in the Changes panel until you approve each one |
| Context / retrieval | Agent-driven search inside the project directory | On-device lexical, symbol, path, outline, and reference search; nothing uploaded to build an index |
| Headless / CI | Terminal-first, so it is scriptable by nature | vortex run and vortex review, with --detach, --json, and --fail-on |
| Remote / SSH work | Runs wherever your terminal runs, as of this writing | SSH remote workspaces; detached jobs survive closing your laptop when started over SSH |
Where Claude Code is the better fit
- You live in the terminal and want the agent there, not behind a GUI.
- You want Anthropic's models only, with one vendor path for the model and the agent.
- You need Linux or Windows today; Vortex's beta is macOS-only.
- You want a single vendor to manage the model, the agent, and support.
- You want the agent to run wherever your terminal already runs, including on a server.
- You prefer a terminal-first tool with no desktop app to install.
- You want to keep the whole loop inside the shell you already script against.
Vortex is macOS-only during the closed beta, and it is not trying to replace a terminal workflow with a window. If the terminal is home and you do not need a desktop surface or third-party keys, Claude Code is the more direct answer.
Where Vortex is the better fit
- A surface you can watch and steer. The macOS app shows the transcript, diffs, the Changes panel, a terminal, sandboxed HTML canvases, and a native WebKit Browser tab.
- Model choice. Bring your own key for OpenAI, Anthropic, OpenRouter, or any OpenAI-compatible endpoint, plus optional local GGUF models via llama.cpp with Metal acceleration.
- A checkpoint before edits land. Review mode keeps agent edits as pending patches in the Changes panel; you approve or reject each one.
- The plan is a file. Plan mode writes
.vortex/plans/<name>.plan.md, so the approach is reviewable and diffable before any edit happens. - Browser handoff. The native Browser tab lets the agent work in a page and hand off to you for CAPTCHA, passkeys, or payment.
- Headless and remote.
vortex runandvortex reviewscript the same loop, and SSH remote workspaces let detached jobs survive closing your laptop. - Rules and memory. Nested
AGENTS.mdand.vortexrulesstack root to leaf, and Cursor-compatible rules work.
What a Vortex job looks like
The same loop runs with or without a window. In the app you watch the transcript, diffs, and Changes panel; headless you script it:
vortex run --prompt "port the parser to the new schema" --workspace . --model openai/gpt-4o-mini
vortex review --json --base main --fail-on errorDetached runs write their status and traces under ~/.vortex/detached/<id>/.
Where the two overlap
Both are job-based agents that read your repository and run commands, so the honest answer is that they overlap a lot. The split is the surface and the wiring: a terminal-first tool tied to one vendor's models, versus a desktop app with BYO keys and a review checkpoint in front of every patch.
A quick way to choose
- If you live in a terminal and want Anthropic's models, Claude Code is the direct answer.
- If you want to watch a job in a GUI and approve patches before they land, Vortex is the direct answer.
- If you need Linux or Windows today, Vortex's beta does not cover it.
- If you want third-party or local models rather than one vendor's, Vortex is built around that.
Sharing a project
Nothing needs to be ported. Both tools read the same repository. Vortex honours AGENTS.md and Cursor-compatible .vortexrules, and MCP servers are configured in Settings → Integrations. Keep whichever tool fits the moment; they do not conflict on disk.
What Vortex does not claim
Vortex is in a closed beta. Verification gates are advisory today: you can require an exit code or an output phrase for a check, but the agent does not hard-block a task on a failing gate. search_memory is lexical, not semantic — there is no embedding corpus or vector search. The beta is macOS 12+ only, so Windows and Linux are out of scope for now. Anything said above about another tool is stated as of this writing; check that tool's own documentation for current behaviour.
Try it
Vortex is invite-only on macOS 12+. Join the waitlist, or compare it with Cursor and read the headless CLI guide before you decide.