AI Engines

Cockpit talks to 6 AI engines out of the box. Each Agent tab picks one engine; you can mix and match across tabs without restarting — pick by what's running locally, what billing account you're on, or which model is best at the task in front of you.

Everything runs locally.

Overview

At a glance

How engine selection works

Each Agent tab has an engine picker in its header. When you create a new tab, the engine defaults to Claude. Switching the engine for an existing tab starts a fresh session — Claude history doesn't carry over into a Codex tab, since each engine has its own conversation format.

You can have, say, six tabs open simultaneously:

• Tab 1: Claude on ~/code/backend
• Tab 2: DeepSeek on the same project for a cheaper second opinion
• Tab 3: Codex on a different project
• Tab 4: Kimi on a notebook, using its long context window
• Tab 5: GLM on a script, billed to your BigModel Coding Plan
• Tab 6: Ollama running a local model for an offline draft

Cockpit's Session Browser (grid icon at the top of the sidebar) shows all of them.

What each engine can do

Image support is engine-level. Ollama tabs silently drop image attachments (no error, but the AI doesn't see them). DeepSeek, GLM and Kimi run the Built-in Agent, which is text-only — see What runs the API-key engines.

What runs the API-key engines

Claude and Codex are driven by their vendor's own tooling — the Claude Agent SDK and the codex CLI. The three API-key engines (DeepSeek, GLM, Kimi) and Ollama all run Cockpit's own Built-in Agent against the provider's OpenAI-compatible endpoint. There is no per-tab mode switch: the engine you pick decides what runs.

What the Built-in Agent gives you: streaming replies, tool calls rendered inline, snapshots, forking, and per-project session files under ~/.cockpit/<engine>-sessions/.

What it doesn't have — the same list for all three engines, so it isn't repeated per section below:

• No images. An images-only message is rejected with "The built-in agent requires a text prompt"; images sent alongside text are dropped and the text is answered.
• Seven tools, no more: Read, Write, Edit, Bash, Glob, Grep, TodoWrite. No subagents, no MCP servers, no web fetch/search.
• No dollar cost readout. Token counts show; the USD total stays at zero. Use the provider's own balance / quota button or console.

Upgrading? Earlier Cockpit versions offered a per-tab Claude Agent SDK / Built-in Agent toggle for these engines. It's gone, and so is the SDK half. Transcripts written by the old SDK mode under ~/.cockpit/<engine>/projects/ are untouched on disk but no longer appear in the sidebar or the Session Browser, since no loop remains that could resume them.

Setting up each engine

Per-engine sections below cover the specifics. Quick pointers:

• Claude — run claude once on your terminal and follow its login prompt. Cockpit reuses your Claude login automatically.
• Codex — install OpenAI's codex CLI and log in with it once. Cockpit reuses that login.
• DeepSeek — get a key from platform.deepseek.com, then paste it in the DeepSeek picker in the Agent tab header (not in the global Cockpit Settings). Pick a model in the same picker.
• GLM — get a key from the BigModel console, then paste it in the GLM picker in the Agent tab header. Pick a model in the same picker, and check the Region row while you're there — see Choose a region.
• Kimi — get a key from the Kimi Code console, then paste it in the Kimi picker in the Agent tab header. Pick a model in the same picker.
• Ollama — install Ollama and pull at least one model (ollama pull llama3.1). When you create an Ollama tab, the model picker lists what you've pulled.

Claude

Claude is Cockpit's default engine — when you start the app and open a new tab, you're talking to Claude unless you pick something else. Cockpit doesn't manage your Claude login; it reuses the claude CLI from Anthropic, so anything you've done there (subscriptions, project settings, MCP servers) is available in Cockpit too.

Setup

You need the Anthropic claude CLI installed and logged in.

1. Install Claude Code if you haven't already:

npm install -g @anthropic-ai/claude-code

2. Log in:

claude

The claude command walks you through the browser-based login. After it's done, Cockpit picks up your credentials automatically — there's nothing to paste into Cockpit.

That's it. Open Cockpit, create a new Agent tab, start chatting.

What you get

• The latest Claude model Anthropic recommends, served through the Claude Agent SDK.
• Image attachments — paste an image into chat (Cmd+V) and Claude can see it. PNG / JPEG / WEBP / GIF up to 5 MB each; you can attach several at once.
• Tool use — Claude can read your files, run shell commands, edit code, hit URLs, use MCP tools.
• Streaming — replies appear word-by-word as Claude thinks.
• Cost visible in the UI — every message shows tokens used and the running USD total per session.

Switching models

Cockpit always uses Anthropic's current recommended Claude model. There is no model picker — you get the latest the service offers. Watch Anthropic's announcements to know which model is current; Cockpit picks it up automatically when the official SDK updates.

Common issues

• "Not logged in" / immediate error on first message — run claude in a terminal and make sure the login completed. Cockpit can only use a login that already works for claude on its own.

Codex

If you have a Codex / ChatGPT subscription, you can drive it from inside Cockpit using the same login. Cockpit doesn't speak the OpenAI API directly here — it spawns OpenAI's codex CLI under the hood and shows you its output.

Setup

1. Install the codex CLI from OpenAI (see OpenAI's docs for the current install command — usually a one-liner).

2. Log in:

codex

Follow the prompt to sign in with your OpenAI account.

3. Open Cockpit, create a new Agent tab, pick Codex in the engine menu. The tab uses your Codex login.

Nothing to paste inside Cockpit — it reuses whatever codex is already configured with on your machine.

What you get

• The Codex model that ships with your CLI (no in-app picker — whatever your codex install gives you).
• Image attachments — Cockpit writes pasted images to a temp file and passes them to codex via the --image flag. PNG / JPEG / WEBP / GIF all work.
• Streaming replies.
• Tool use — Codex can read your files, run shell commands, and edit code.
• Multi-tab sessions — open as many Codex tabs as you need, each independent.

What you don't get

• No running cost display. Cockpit can't read pricing back from the codex CLI, so the token bar stays empty for Codex tabs (total_cost_usd: 0). Track usage on your OpenAI dashboard instead.
• No model picker. Whichever model your codex CLI uses is what runs.

Common issues

• "codex not found" / nothing happens on send — the codex CLI isn't on your PATH. Verify with codex --version in a terminal; if that fails, reinstall.
• Login expired — re-run codex in a terminal and complete the login flow again. Cockpit doesn't manage the login itself.
• Outdated CLI — OpenAI updates codex periodically. If something behaves oddly, upgrade.

DeepSeek

DeepSeek is the cheapest cloud engine in Cockpit. Unlike Claude / Codex (which reuse a CLI's login), DeepSeek is API-key-only — paste a key in the per-tab DeepSeek picker and you're done. GLM and Kimi work the same way.

Under the hood it goes through DeepSeek's OpenAI-compatible endpoint (https://api.deepseek.com/v1), driven by Cockpit's Built-in Agent — so tool calls and streaming work, but images, MCP and subagents don't.

Setup

1. Get an API key from platform.deepseek.com. It looks like sk-....

2. Open a new tab in Cockpit, pick DeepSeek in the engine menu, then click the DeepSeek picker icon in the tab header → paste the key into the API Key field → save. (The key lives in its own credential file, ~/.cockpit/deepseek/credentials.json — not in ~/.cockpit/settings.json, and not in the global Cockpit Settings modal. settings.json only remembers which model you picked.)

3. Pick a model in the same picker.

Done. The key only ever stays on your machine.

Pick a DeepSeek model

The model list is fetched live from DeepSeek's GET /v1/models using your key — Cockpit whitelists nothing, so a model DeepSeek ships tomorrow shows up on its own (and an id your account can't serve is rejected by DeepSeek with its own message).

A new DeepSeek tab defaults to deepseek-v4-pro. The flash variants are the cheap, fast end of the list — good for quick fixes, formatting and simple Q&A; pro is slower and smarter, for architecture decisions, hard bugs and multi-step refactors.

Check your balance

DeepSeek is prepaid, so the DeepSeek tab gets a balance button next to the model picker. It needs a saved key and doesn't poll — it fetches only when you click.

What you get

• A model picker, live from DeepSeek's API (see above).
• Streaming replies.
• Tool use — DeepSeek can read your files, run shell commands, edit code.
• Forking and per-tool-call snapshots, like every other engine.
• Token counts visible in the UI.

What you don't get

Everything in What runs the API-key engines applies: no images, seven tools only, and no dollar total — the token bar's USD figure stays at zero. Use the balance button, or check your DeepSeek dashboard for the real bill.

Common issues

• "DeepSeek API key is not configured" — you haven't pasted a key in the picker. The key goes in the per-tab DeepSeek picker in the header, not in the global Cockpit Settings modal.
• "401 / Unauthorized" — bad or expired key; paste it again in the picker and watch for stray whitespace.
• Slow / hanging replies — pro is genuinely slower than flash; if you don't actually need the reasoning, switch the tab to a flash model.
• Spending climbing fast — pro is several times more expensive than flash. Click the balance button to see where you stand.

GLM

GLM is Zhipu AI's model family, sold through the BigModel platform. Structurally it's the same kind of engine as DeepSeek and Kimi: API-key-only, with a live model picker, a quota readout, and a fork-able session store.

The one thing GLM has that no other engine does: it's served from two hosts — one in mainland China, one international — and the picker has a Region row to pick between them. Same key either way; see Choose a region.

Setup

1. Get an API key from the BigModel console. GLM keys are two dot-separated halves, <id>.<secret> — no sk- prefix.

2. Open a new tab in Cockpit, pick GLM in the engine menu, then click the model picker in the tab header → paste the key into the API Key field → save. (The key lives in its own credential file, ~/.cockpit/glm/credentials.json — not in ~/.cockpit/settings.json, and not in the global Cockpit Settings modal.)

3. Pick a model in the same picker. Cockpit fetches the list from your account the moment the key is saved.

Done. The key only ever stays on your machine.

Pick a model

The model list is fetched live from GLM's GET /models using your key — Cockpit hard-codes nothing, so new models show up on their own. At the time of writing the API returns eight:

glm-4.5 · glm-4.5-air · glm-4.6 · glm-4.7 · glm-5 · glm-5-turbo · glm-5.1 · glm-5.2

A new GLM tab defaults to glm-5.2.

GLM publishes no per-model metadata. Its model list is bare ids — no display name, no context window. So the picker shows ids only, and Cockpit sets no context-window hint for GLM tabs (Kimi tabs get one because Kimi reports it). That's a limit of the provider's API rather than a missing Cockpit feature.

glm-5.2[1m]. BigModel's Claude Code docs mention that [1m] suffix for a 1M-token context. Cockpit used to reject it outright — its old Anthropic-compatible route answered HTTP 400 "模型不存在". GLM tabs now go through the OpenAI-compatible coding endpoint instead, so if you need the long-context variant, try it and let the provider answer; a model id your account can't serve comes back as GLM's own error.

Choose a region

GLM is served from two hosts, and the same key works on both: a key issued on bigmodel.cn authenticates on z.ai and reports the identical quota. The region is pure routing.

The default comes from your Cockpit UI language: English → International, everything else (including "auto") → mainland. Override it in the Region row of the GLM picker — your pick wins and is remembered.

Language only seeds that default; it never overrides a choice you made. That's deliberate: changing Cockpit's UI language shouldn't silently re-route your API traffic to a server in another country. If your language and your account don't line up, set the region once and forget it.

Sessions are not region-scoped. Switch regions whenever you like — existing GLM conversations stay resumable, and the key stays valid.

Check your Coding Plan quota

GLM's Coding Plan is a subscription, so the GLM tab gets a Check quota button next to the model picker rather than a dollar total. Click it and Cockpit reads what's left in two windows:

• a rolling 5-hour window, and
• a weekly one.

Each shows as remaining/limit, prefixed by your plan tier — e.g. lite · 5h 1990/2000 · 1w 4980/5000. Hover for when the longer window resets; it goes red when a window is exhausted. The button needs a saved key and doesn't poll — it fetches only when you click.

No Coding Plan means no quota, and that's normal. A plain pay-as-you-go BigModel key has no plan allowance to report, so the button answers "Quota unavailable — check the API key". Chatting still works exactly as before — you're just billed per token instead. The link next to the button opens BigModel's usage page for the real figures.

What you get

• A model picker, live from GLM's API (see above).
• A region switch — mainland or international, same key, sessions unaffected.
• Streaming replies.
• Tool use — GLM can read your files, run shell commands, edit code.
• Forking — fork a GLM session from any message, same as Claude.
• Per-tool-call snapshots, like every other engine.
• Multi-tab sessions, each independent.

What you don't get

Everything in What runs the API-key engines applies — no images, seven tools, no USD total. GLM-specific on top of that:

• No dollar cost readout at all. The token bar's USD figure stays at zero. On a Coding Plan use Check quota; on pay-as-you-go check the BigModel console.
• No context window or display name in the picker — GLM doesn't report either. See the note under Pick a model.

Common issues

• The picker says "Set API key" — no key saved yet. It goes in the GLM picker in the tab header, not the global Cockpit Settings modal.
• "Failed to load models — check the API key" / 401 — bad or expired key. Re-paste it from the BigModel console and watch for stray whitespace; a GLM key is the full <id>.<secret> string, both halves included.
• HTTP 400 "模型不存在" — that model id isn't served to your account. Pick one from the list the picker fetched.
• Quota says unavailable but chat works — your account has no Coding Plan. Nothing is broken; see Check your Coding Plan quota.
• Slow or flaky connection — you may be on the far host. Flip the Region row: mainland accounts are usually fastest on open.bigmodel.cn, and the same key works on api.z.ai if you're outside China. Switching is safe; sessions survive it.
• Pasted images are ignored — GLM tabs are text-only. Use a Claude or Codex tab when you need the model to see an image.

Kimi

Kimi is a Chinese-market AI from Moonshot, known for long context windows. Cockpit talks to Kimi Code directly over its API — paste a key in the per-tab Kimi picker and you're done. Structurally it's the same kind of engine as DeepSeek: API-key-only, with a model picker and a fork-able session store.

Changed: Cockpit no longer uses Moonshot's kimi CLI. You don't need it installed, and its login no longer matters. See Upgrading from the kimi CLI if you used Kimi in an earlier Cockpit.

Setup

1. Get an API key from the Kimi Code console. It looks like sk-kimi-....

   This is a Kimi Code key, not a Kimi Open Platform key from platform.moonshot.cn. The two are separate products and the keys are not interchangeable — an Open Platform key will fail here.

2. Open a new tab in Cockpit, pick Kimi in the engine menu, then click the model picker in the tab header → paste the key into the API Key field → save. (The key lives in its own credential file, ~/.cockpit/kimi/credentials.json — not in ~/.cockpit/settings.json, and not in the global Cockpit Settings modal.)

3. Pick a model in the same picker. Cockpit fetches the list from your account the moment the key is saved.

Done. The key only ever stays on your machine.

Pick a model

The model list is fetched live from Kimi's GET /coding/v1/models using your key — it isn't a list Cockpit hard-codes. Which models come back depends on your membership tier, so two accounts can see different lists, and the list changes when Kimi ships or retires a model.

At the time of writing a Kimi Code account can see:

Don't treat that table as fixed — if a model isn't in your picker, your plan doesn't include it. The picker shows the model id, its display name when it differs, and its context window.

Check your quota

Kimi Code is a subscription, not a prepaid balance, so the Kimi tab has a Check quota button next to the model picker instead of a cost readout. Click it and Cockpit reads what's left in:

• your plan cycle — a 7-day window, and
• a rolling 5-hour window on top of it.

Both show as remaining/limit (e.g. plan 100/100 · 5h 40/50); hover for when the plan window resets. It goes red when a window is exhausted. The button needs a saved key and doesn't poll — it only fetches when you click.

What you get

• A model picker, live from your account (see above).
• Streaming replies, with the model's "thinking" steps folded into a <details> block before the final answer.
• Tool use — Kimi can read your files, run shell commands, edit code.
• Forking — fork a Kimi session from any message, same as Claude. (This didn't work when Kimi went through the CLI.)
• Per-tool-call snapshots, like every other engine.
• Multi-tab sessions, each independent.

What you don't get

Everything in What runs the API-key engines applies — no images, seven tools, no USD total. Kimi-specific on top of that:

• No dollar cost readout. Kimi Code bills by subscription, so there's nothing per-token to total up — use Check quota instead.

Upgrading from the kimi CLI

If you used Kimi in an earlier Cockpit, two things changed and both are breaking:

• Your kimi CLI login no longer applies. Get a Kimi Code key and paste it into the picker. The CLI itself is no longer used or required — you can uninstall it as far as Cockpit is concerned.
• Old Kimi transcripts are gone from Cockpit. Sessions used to live in ~/.kimi, and later in ~/.cockpit/kimi/projects/ (the removed SDK mode); Cockpit indexes neither, so those conversations won't appear in the sidebar or the Session Browser. The files are untouched on disk if you want to read or archive them yourself. New sessions land under ~/.cockpit/kimi-sessions/.

Common issues

• The picker says "Set API key" — no key saved yet. It goes in the Kimi picker in the tab header, not the global Cockpit Settings modal.
• "Failed to load models — check the API key" / 401 — bad, expired, or wrong-product key. Re-paste it (watch for stray whitespace) and confirm it's a sk-kimi-... key from the Kimi Code console rather than a platform.moonshot.cn one.
• A model you expected isn't listed — the list is gated by membership tier. k3 needs Moderato or higher, kimi-for-coding-highspeed needs Allegretto or higher.
• Replies stop mid-day — you may have hit the 5-hour window. Click Check quota; if it's red, wait for the reset shown in the tooltip.
• Pasted images are ignored — Kimi tabs are text-only. Use a Claude or Codex tab when you need the model to see an image.

Ollama

Ollama is the only engine in Cockpit that runs entirely on your own machine. No API key, no cloud, no per-token cost. Install Ollama, pull the models you want, and Cockpit lists them in the model picker.

Reach for this engine when:

• You're on a plane or otherwise offline.
• You're working with sensitive code that shouldn't leave your laptop.
• You have a workstation with a beefy GPU and want to use it.
• You're experimenting with custom or fine-tuned models.

Setup

1. Install Ollama from ollama.com.

2. Pull at least one model:

ollama pull llama3.1

You can pull more later: ollama pull qwen3.5, ollama pull deepseek-coder, etc. See the Ollama model library for the full list.

3. In Cockpit, create a new Agent tab and pick Ollama in the engine menu. If the Ollama service isn't running, Cockpit auto-spawns ollama serve and waits up to 8 seconds for it to be ready.

4. Click the model dropdown in the tab header — Cockpit asks the Ollama API for the list of models you've pulled.

What you get

• Any model you've pulled, picked per tab.
• Streaming replies.
• Tool use (depends on the model — coding-tuned models support tool use, generic chat models often don't).
• Completely offline. No outbound network calls.
• Zero per-message cost.

What you don't get

• No image attachments. Cockpit's Ollama tab is text-only for now, even if you pull a vision-capable model.
• No "best practice" model picker. Ollama gives you exactly what you pulled — Cockpit has no opinion. (There's a fallback default in code, but you should pick one yourself.) Start with a known-good coding model like qwen3.5-coder or deepseek-coder if you're unsure.

Choosing a model

Rough sizing guidance — actual performance varies by GPU:

For coding work specifically, look at qwen3-coder, deepseek-coder, and codellama families. They're more useful than a generic chat model of the same size.

Common issues

• "No models found" in the dropdown — you haven't pulled anything yet. Open a terminal and run ollama pull <name> for at least one model.
• Replies are extremely slow — the model is bigger than your GPU can comfortably handle. Try a smaller one.
• Auto-start didn't work — run ollama serve in a terminal manually and try again.