Kingcurve user guide

An AI coding assistant for VS Code. Use local models (Ollama, LM Studio, any OpenAI-compatible server) or your own cloud API keys (OpenAI, Anthropic Claude, Google Gemini, OpenRouter). Kingcurve has no server of its own: requests go straight from VS Code to the provider you choose.

• Chat about your code with streaming answers, syntax highlighting and copy buttons.
• Attach context with @file, @folder, @selection, @search, @problems, @git, @terminal and @workspace.
• Edit multiple files with a diff review: you choose which files and changes to keep, and you can roll back.
• Agent mode reads files, edits, runs commands and tests, and checks its own work, asking for your approval along the way.
• MCP servers: connect Model Context Protocol tools, resources and prompts.
• History, token usage and cost, commit messages from your staged changes, and optional inline autocomplete.

Getting started

1. Install Kingcurve. It opens in the Secondary Side Bar on the right, next to other AI chats. Press Cmd+Alt+M (macOS) / Ctrl+Alt+M (Windows, Linux) to open it, or toggle the Secondary Side Bar with Cmd+Alt+B / Ctrl+Alt+B. You can drag it to another place, such as the Activity Bar on the left.
2. Connect a model (see below). Ollama and LM Studio running on your machine are detected automatically.
3. Pick a mode and a model in the message box (bottom left) and start asking. Use Add context or type @ to attach files, and paste images directly.

Requires VS Code 1.106 or newer.

Connecting a model

Open the Providers tab in the sidebar, or run Kingcurve: Manage Providers. You can add as many providers as you like, including several of the same type. Models are discovered from each provider; you can also type a model name.

Ollama

1. Install Ollama and pull a model, for example ollama pull qwen2.5-coder.
2. The "Ollama" provider appears automatically (http://127.0.0.1:11434/v1). Click Test, then choose a model.

The context length is adjusted automatically. Kingcurve starts at 8K tokens and grows the context (num_ctx) as the conversation gets longer: 16K, then 32K. Agent mode starts at the maximum, because it reads many files. The maximum is 32K by default (Settings → Maximum automatic context), and never more than the model supports. To use a fixed value instead, set it in the provider's advanced options or in Settings.

LM Studio

1. In LM Studio, open the Developer tab and start the server (port 1234).
2. The "LM Studio" provider appears automatically. Click Test and choose a model.

The context length is adjusted automatically. LM Studio sets a model's context when it loads it, often to only 4K tokens. When a conversation or Agent mode needs more, Kingcurve reloads the model with a larger context through LM Studio's API (up to the automatic maximum, 32K by default). The message box shows when this is happening. This needs LM Studio 0.4 or newer; with older versions, load the model with a larger Context Length yourself.

LM Studio on another computer: set the provider's Base URL to that machine, for example http://192.168.1.42:1234/v1, and turn on Serve on Local Network in LM Studio. If LM Studio has Require Authentication on, paste its API token into the API key field. Then tick Allow API key over http on the local network, because LM Studio uses plain http. Only private network addresses are accepted. Anyone on the same network could read the token, so use this only on a network you trust.

Cloud providers

Click Add provider, choose OpenAI, Anthropic Claude, Google Gemini or OpenRouter, paste your API key and save. Usage is billed by that provider to your account.

Other OpenAI-compatible servers

Choose OpenAI-compatible and enter the base URL that serves /chat/completions (usually ending in /v1). This works with vLLM, the llama.cpp server, LocalAI and hosted gateways.

Adding context

Type @ in the message box, or click @, to attach files, folders, your selection, search results, problems, git changes or terminal output. You can also write references inline:

Why does @file:src/cart.ts#L10-40 return NaN? See @problems and @git.
Where else is @search:checkoutTotal used?

Chips under each message show what was attached and how many tokens it used. Large content is compressed or trimmed to fit the model. Kingcurve respects .gitignore and .mycodeignore, and never reads .env files, private keys or other credential files. Secret-looking values are hidden before anything is sent.

You can also right-click a selection or a file and choose Add Selection to Chat or Add File to Chat.

Images and files

Attach screenshots, mockups and design images (PNG, JPEG, WebP, GIF) or text files (code, Markdown, JSON, CSV, …) from anywhere on your computer:

• click the 📎 button next to the message box,
• paste an image (for example a screenshot) with Cmd+V / Ctrl+V, or
• drag files onto the message box.

For example, attach a UI mockup and write "Build this page in Vue with Tailwind, following this layout". The model sees the image and can propose the code, which you review as diffs as usual.

• Images need a model with vision: GPT-4o or newer, Claude, Gemini, or a local vision model (for example llava, qwen2.5vl or gemma3 in Ollama, or a VLM in LM Studio). Kingcurve warns you when the selected model cannot read images.
• Large images are scaled down to 1568 px on the long edge before sending. Up to 6 attachments per message; text files up to 200,000 characters.
• Images stay available for follow-up questions in the same conversation, including after a restart. When the conversation gets long, older images are left out first.
• Attached text files go through the same protections as workspace files: credential files (.env, keys) are refused, and secret-looking values are hidden.
• SVG and PDF are not supported yet: attach a PNG export or a screenshot instead.

Editing code

Ask for a change, or select code and run Kingcurve: Edit Selection with AI…. A review card shows every proposed change; open any file in the VS Code diff editor and keep or drop individual files and change blocks.

• Nothing is written until you click Apply.
• Files you changed in the meantime are never overwritten.
• Roll back restores the previous content, even after restarting VS Code.
• Files that can run code (build scripts, CI, git hooks, package.json) need an extra confirmation. .git/, node_modules/ and credential files are never changed.

Modes and the agent

Choose a mode under the message box:

Agent mode needs a model that supports tool calling; Kingcurve tells you if the selected model does not. You stay in control:

• Every file change goes through the diff review.
• Every command asks for approval by default. Destructive commands are blocked, and installs or network commands always ask.
• Commands are shown in a Kingcurve Agent terminal.
• Stop cancels everything at once.
• Show Agent Audit Log lists every action the agent took.

You can relax these defaults in Settings, for example to apply edits automatically or to run commands classified as safe without asking.

MCP servers

The MCP tab connects Model Context Protocol servers: local programs (stdio) or remote servers (Streamable HTTP). You can paste or import the usual mcpServers JSON.

• Server tools become available to the agent. Each tool can be set to allow, ask (the default) or deny.
• Resources can be attached to a message, and prompts inserted into the message box.
• A local server starts only after you approve its exact command. It gets a minimal environment.
• Remote servers must use https.

History, usage, git and autocomplete

• History: conversations are saved on your computer and the last one reopens with VS Code. You can search, rename, export (Markdown or JSON) and delete them.
• Usage: tokens and estimated cost by day, model and feature, plus a counter in the status bar. Cost is shown when the provider reports prices (OpenRouter) or when you add prices in mycode.usage.modelPricing.
• Commit messages: the ✨ button in Source Control writes a message for your staged changes into the commit box for you to edit. Nothing is committed for you.
• Review Uncommitted Changes and Explain Last Terminal Command send your diff or your last command's output to chat.
• Autocomplete (off by default): inline suggestions from a model you choose. Ollama coder models use fast fill-in-the-middle. Autocomplete never runs on sensitive files.

Privacy

• Kingcurve has no backend and collects no telemetry.
• Your messages, and the context you attach or the agent reads, are sent only to the model provider you selected. MCP tool calls go only to that MCP server. With a local model, nothing leaves your computer.
• API keys and tokens are stored in your operating system's keychain (VS Code SecretStorage), never in settings files.
• Conversation history is stored locally in VS Code's storage for the extension, readable only by your user account.

Security and untrusted folders

• In folders you have not marked as trusted (Restricted Mode), Kingcurve offers chat only. It does not read files, edit, run commands, start MCP servers or autocomplete until you trust the folder.
• Kingcurve's settings can only be set in your user settings. A repository cannot change them through .vscode/settings.json.
• Model output is treated as untrusted: it cannot load remote images or run scripts in the chat view, and links open only through VS Code's own link confirmation.
• Attached files and tool results are marked as data for the model, which reduces (but cannot fully prevent) prompt injection. Always review diffs and commands before approving them.

Commands

In the message box: Enter sends, Shift+Enter adds a new line, Esc stops the reply.

Settings

Open the Settings tab in the sidebar, or search for mycode in VS Code settings. You can change:

• generation (temperature, maximum output, context length, extra instructions);
• timeouts and retries;
• workspace context (token budget, exclusions, indexing);
• file editing;
• agent limits and the read, write and terminal permissions;
• autocomplete, history retention and model prices.

Known limitations

• Token counts are estimates for providers that do not report usage. Estimates are marked "≈".
• Agent mode depends on the model: small local models often call tools unreliably.
• Remote MCP servers that require OAuth sign-in are not supported yet. Use a token instead.

License

Kingcurve is free software under the MIT License. Third-party components it bundles are listed in THIRD_PARTY_NOTICES.md.