Codepit

A graphical cockpit for coding agents inside VS Code. Claude Code is the default, with Codex, Kimi Code and local models through Ollama as options. Chat in a native panel, keep an eye on your plan limits, review and undo what each turn changed, and pick up Shortcut stories — while each agent keeps owning its sign-in, MCP servers, skills, hooks and permissions.

The extension never asks for an API key for Claude Code, Codex or local models: it drives the CLIs you already have installed and signed in, and local models run on your own computer. The optional experimental Ahefi provider and the optional extra Claude accounts are the exceptions; their keys are kept in VS Code's encrypted SecretStorage.

Features

Chat panel

Open it from the Codepit icon in the activity bar or with Codepit: Open. Choose the model, the effort (low to max) and the permission mode (default, accept edits, plan, auto or bypass) per chat, follow every tool Claude runs, and queue a message while a turn is still running. A context meter tells you when to compact the conversation and does it with one click. Every turn lists the files it edited, and you can undo the whole turn.

Review each turn

Review opens one tab with a unified or side-by-side diff of everything that turn changed. Files brought in by a git checkout or pull during the turn are kept apart from Claude's own edits.

Permissions and questions

A permission prompt says what the tool is about to do (the command, the file and the change, the web address…) and keeps the raw input under Technical details. Claude's questions open as a form, and your answers reach Claude.

Effort estimates for each session

Click Session effort in the header to open a floating task table with estimated manual hours and an optional AI-assisted column. Enable it in the dialog's project settings, write your own estimation prompt. Find my criteria in memory searches the evaluator's memory and instructions for this project (Claude's project memory and CLAUDE.md, Codex's memories and AGENTS.md) for passages about estimates, and the evaluator turns only those passages into a prompt you review before saving; you can also import a Markdown/text file. Any provider you have linked can evaluate (Claude, Codex, Kimi Code, a local Ollama model or Ahefi), with a model from its own list. Codepit recommends one: a fast, inexpensive model from a provider that guarantees the table format, Claude Haiku first. The evaluator does not have to be the chat's provider and stays the same for every session of the project, so estimates compare. Evaluations consume that provider's tokens (a local model is free) and run separately from your chat.

The header keeps a live preview, with automatic updates after turns and during longer work when new evidence is available. Each saved estimate keeps its criteria and update time. These are effort ranges based on session evidence, rather than measured human working hours.

Plan limits at a glance

The header shows the 5-hour and weekly windows of your Claude plan, including per-model weekly limits and extra-usage credits, with charts of where the usage went and what this session used.

Keep going when an account runs out

Register more Claude accounts in Settings → Claude accounts: sign in to another subscription (Pro, Max, Team or Enterprise) or to the Claude Platform (billed per token) on Claude's own sign-in page, as /login does. Nothing is copied or pasted. When the account in use reaches its 5-hour or weekly limit, the chat shows the other accounts with their usage, or lets you add one right there; pick one and the conversation resumes there with all its context. The header always shows the plan you are working on (Team 5x, Max 20x, Pro or API), switches account in one click and adds another from the same menu; the limits dialog follows the account in use. On a Claude Platform account it shows what its chats spent today and this month and, with the balance you enter from the Console (Anthropic offers no way for apps to read it), how much is left. Your Claude Code settings, skills, MCP servers and session history stay the same on every account, and Codepit offers to switch back when the first account resets.

Lighter sessions for projects with a long CLAUDE.md

Claude Code sends the whole CLAUDE.md at the start of every session. When yours weighs over 30 KB, the header shows Optimize CLAUDE.md with its size. It opens a chat that runs the bundled slim-claude-md skill: the sections Claude only needs for some tasks move, word for word, into files Claude Code loads on demand (rules for their folders, skills, docs). You see the plan first, nothing changes until you approve it, a backup is kept and the skill checks that no line was lost. Codepit installs the skill in your Claude Code skills folder the first time, after telling you.

Sub-agents

When Claude spreads the work over agents, the agent map shows each one live: what it is doing, its tools and tokens, nested and background agents, and the report of the ones that fail.

History

Search, filter and organize chats, pin the history beside the chat, and import sessions you started in Claude Code. Export one chat, a selection or the project as JSON or Markdown, preview a JSON backup and choose which conversations to restore. Optional local backups keep the latest ten copies.

Shortcut stories

Paste a read-only Shortcut API token once per project to list your stories. Review or Code opens a chat for the story, and reopens the same chat the next time.

Codex (optional)

Turn on ChatGPT models through the Codex CLI as a second provider from Settings → Providers, then pick the provider and the model for each new chat. It is off by default.

Kimi Code (optional)

Turn on Kimi Code from Settings → Providers. Codepit can install its CLI, open its sign-in flow, list available models and run each chat as a resumable Kimi session. On Windows, Git for Windows is required by Kimi Code.

Local models through Ollama (beta)

Work on your project with a model running on your own computer, so nothing leaves it. Any model Ollama has that can use tools shows in the model picker (Qwen, Llama, Mistral…). The agent is Qwen Code, with Codepit's permission modes, undo and review.

• Settings → Providers → Local (Ollama) → Set up installs whatever is missing — Node, the Qwen Code engine, Ollama and a model — with each step and its progress in view. Node and the engine go into Codepit's own folder, so your Node and npm stay as they are.
• The model is chosen for your GPU: Qwen3.8 27B from 24 GB of VRAM, Qwen3.5 9B or 4B below that.
• Ollama does not start with your computer: Codepit starts it when a chat needs it.
• Uninstall… lists what local models use and removes what Codepit installed; anything you installed yourself is listed too, unticked.

This provider is in beta: its installer has not yet been tested on a clean Windows, macOS or Linux machine.

And also

• Clickable files — file paths and links in Claude's replies open the file at the right line.
• Speed per turn — the last reply of each turn shows tokens per second and the wait for the first token.
• Context in use — the session meter shows how much of the model's context window the last call filled.
• Attachments — attach files and images with the paperclip, Ctrl+V or drag and drop.
• Ahefi (experimental) — optional direct HTTPS chat using an Ahefi API key. It is isolated, off by default and chat-only: it receives no workspace files, tools, terminal or MCP access.
• Pinned side panels — pin history, Shortcut and settings next to the chat.
• English and Spanish, following your choice in the Codepit settings, and light or dark, following your VS Code theme.

Requirements

• VS Code 1.95 or newer.
• Claude Code installed and signed in once (claude, then /login). Codepit's Settings → Providers section can install it and open the sign-in for you. The Claude Code VS Code extension is not needed.
• Optional: the Codex CLI signed in with codex login, only if you want ChatGPT models.
• Optional: a GPU for local models; Set up installs the rest.
• Optional: an Ahefi API key, only if you explicitly enable the experimental chat-only provider.

Codepit is developed and tested on Windows. It also looks for the CLIs in the usual macOS and Linux locations.

Settings

Privacy

• Claude and Codex chats run through their local CLIs, under each CLI's own account and settings.
• Local chats never leave the computer. Set up downloads Node, Qwen Code, Ollama and the model only from their official sources, and only when you press it.
• Ahefi chats are sent directly to https://ahefi.com/v1/chat/completions. The API key is stored in VS Code SecretStorage, never in settings or chat history, and is not written to the output log.
• To show your plan limits, the extension reads the Claude Code sign-in stored on your computer and asks Anthropic's usage service for your current usage. It keeps only the last reading.
• Extra Claude accounts: Claude Code itself signs in to each extra account and keeps that sign-in (or the API key a Claude Platform sign-in creates) in a folder of Codepit's own storage, apart from your main one; removing the account signs it out. Only the Claude Code processes of that account use it, and Codepit sends a subscription's sign-in to Anthropic alone, to read its usage.
• Shortcut tokens are kept in VS Code's secret storage and are sent only to Shortcut's API.
• History exports include chat text and metadata, so review the file before sharing it. Codepit excludes tool payloads and stored credentials and redacts recognized secret patterns.
• The extension sends no telemetry.

Updates

Updates arrive automatically through the Visual Studio Marketplace. VS Code installs a new version in the background and asks you to restart extensions; your chats and settings are kept.

Support Codepit

If Codepit saves you time, you can buy me a coffee ☕.

Codepit is made by Vicxelcode, S.L. It is not affiliated with, endorsed by or sponsored by Anthropic, OpenAI, Alibaba (Qwen), Ollama or Shortcut. Claude and Claude Code are trademarks of Anthropic.