Use DeepSeek inside Cockpit — and keep all your Claude habits

If you already use Cockpit with Claude, you have a workflow: open a tab, ask the agent to fix a bug, watch it edit files, run tests, hand you a clean diff. Slash commands like /qa and /fx are muscle memory.

Now you can do all of that with DeepSeek instead — usually at a fraction of the cost. As of v1.0.195, DeepSeek sits next to Claude in the new-tab menu, and everything you already know how to do works exactly the same.

Setup: under a minute

1. Get a key. Go to platform.deepseek.com/api_keys, create one, copy it.
2. Open a Deepseek tab. Click the + on the tab bar, pick Deepseek.
3. Paste the key. A blue Set API key pill appears in the chat header. Click it, paste, hit Save.
4. Send a message. That's it.

Your key stays on your laptop, in ~/.cockpit/settings.json. Cockpit has no server — nothing leaves your machine except the request to DeepSeek itself.

Pick a model from the same pill

Once the key is in, that pill turns into the current model name. Click it again to switch:

• deepseek-v4-pro — the default. Use it for the kind of tasks you'd give Claude Sonnet: refactors, debugging, multi-file edits, writing tests.
• deepseek-v4-flash — faster and cheaper. Great for "do this small thing" tasks: rename a function, write a one-off script, summarize a file.

You can have one tab on pro and another on flash at the same time. Each tab remembers its own model.

What it can do (spoiler: everything)

The whole point of plugging DeepSeek into Cockpit is that nothing else changes. In a Deepseek tab the agent can still:

• Read and edit files in your project.
• Run terminal commands — install deps, run tests, start a dev server.
• Search the codebase with Grep / Glob.
• Browse the web — WebFetch and WebSearch work the same.
• Run your slash commands — /qa to clarify before coding, /fx to diagnose a bug, /review for a pre-PR look, /commit to wrap up.
• Spawn sub-agents for parallel research / refactor work.
• Take screenshots / pasted images as input.
• Pick up where you left off — close the tab, reopen tomorrow, the conversation is right there.

If you've built habits around Cockpit's Claude experience, your habits transfer over wholesale. Same buttons, same shortcuts, same flow.

A few real workflows worth trying

The "cheap second opinion". Open two tabs side-by-side: one Claude, one DeepSeek-v4-pro. Paste the same prompt into both. Compare answers. You'll learn fast which model your codebase prefers — and the comparison itself usually surfaces a better question.

Bulk grunt work on Flash. That afternoon of "rename this prop across 40 files, update the storybook stories, regenerate types" — point a Flash tab at it. It's plenty smart for mechanical changes, and noticeably faster.

/fx on DeepSeek. Bug-evidence mode (/fx The login modal sometimes flashes empty) works particularly well here — the agent reads the failing path, builds a hypothesis, and stays out of your code until you say go.

Long sessions without the bill anxiety. Long agent conversations on Claude can get expensive once the context grows. DeepSeek's pricing means you can let a session breathe — keep iterating, keep showing it more files, keep refining — without watching the meter.

Things kept clean and separated

Two practical promises about how DeepSeek lives next to Claude on your machine:

• Conversations don't mix. A DeepSeek session is stored separately from your Claude history. Searching Claude history won't surface DeepSeek replies, and vice versa.
• Credentials don't leak. Your Claude login and your DeepSeek key live in different places. One has nothing to do with the other.

You can swap between Claude and DeepSeek tabs all day and never worry about cross-contamination.

How to start right now

npm i -g @surething/cockpit
cock

Open the app, pick Deepseek from the new-tab menu, paste your key, ask it to fix something. If it understands your codebase as well as Claude does, the cost per task may surprise you.

npm i -g @surething/cockpit · GitHub · Try Online