Kova

Kova is a local-first AI coding agent for VS Code. It keeps Ollama as the default offline provider, while also supporting DeepSeek and configurable OpenAI-compatible APIs when you want a remote model.

Kova opens in the Activity Bar sidebar, with an optional editor tab. It can answer questions, inspect workspace context, edit files and run checked commands through the same permission and approval pipeline regardless of the selected model provider.

Providers

API keys are stored with VS Code SecretStorage, not in settings.json, workspace files or the repository.

Get started

Local demo with Qwen3 4B

1. Install and start Ollama.
2. Install the default model: ollama pull qwen3:4b.
3. Install Kova from the Marketplace or from the generated VSIX.
4. Run Kova: Open Chat from the Command Palette.
5. Keep kova.provider set to ollama, select qwen3:4b, then choose a mode.

DeepSeek or another API provider

1. Open Kova: Open Settings and change kova.provider to deepseek or openaiCompatible.
2. Configure the provider endpoint/model if needed.
3. Run Kova: Set Provider API Key when the endpoint requires authentication.
4. Use Kova: Clear Provider API Key to remove the stored credential.

Open Kova from the Activity Bar or run Kova: Open Chat. For a movable tab alongside your code, run Kova: Open Chat in Editor. The sidebar and editor tab share the same conversation. Closing the editor tab stops active work; hiding the sidebar keeps the run active.

Privacy and network behavior

Kova does not silently switch providers. ollama is the default and sends requests only to the configured Ollama endpoint, which defaults to http://127.0.0.1:11434.

When you explicitly select DeepSeek or an OpenAI-compatible provider, prompts, selected conversation context, tool definitions and tool results needed for the active request are sent to that configured API endpoint. Provider credentials are read from VS Code SecretStorage and are not added to model context, logs or workspace configuration.

Thinking/reasoning is displayed separately when supported and is not persisted in Kova conversation history. DeepSeek reasoning required for a thinking + tool-call continuation is kept only as transient in-memory provider state for that conversation.

Modes

Protected configuration writes require approval in every mode. Destructive commands are blocked. Approvals apply once to the prepared operation; a file changed after preview requires a new preview.

Workspace tools

Read files, list directories, search text, inspect Git diffs, create files, replace one exact text match and run checked commands. Built-in file tools stay inside the selected workspace. Changes support VS Code Undo.

Use Add context to attach a selection, the current file or chosen files. When chat has focus, selection and current-file attachments use the most recently active file editor. Kova does not index the workspace in the background.

Stop cancels generation, pending approval or a running process. Applied changes remain. New conversation clears the chat. Conversation history lasts for the current extension session.

Context and settings

The default context budget is 32,768 tokens, with 4,096 reserved for output and a 10% safety margin. The context indicator shows estimated input before generation and measured input/output afterwards. Tool-definition cost remains an estimate. Larger local contexts require more memory; explicit VS Code settings override these defaults.

Run Kova: Open Settings to change:

Generic OpenAI-compatible thinking is disabled because reasoning wire formats are not standardized. Auto allowlisting matches the whole command; chaining, pipes, redirection, substitutions and backticks never match.

Skills, MCP and hooks

Select a skill from .kova/skills/<id>/SKILL.md; only its selected body enters context. Configure local stdio MCP servers in .kova/mcp.json and command hooks in .kova/hooks.json. MCP tools use the same permission and guardrail pipeline as built-in tools. Configuration changes take effect between runs.

See the usage guide for skill and MCP examples, thinking display and permission modes.

Use trusted projects. Configured MCP servers, hooks and project scripts run with your user privileges. They are not sandboxed. Kova adds no separate workspace-trust prompt. Writes to .kova/mcp.json, .kova/hooks.json, .git/** and .vscode/** require approval.

Requirements

• VS Code 1.100 or newer.
• For the default local mode: Ollama and an installed model such as Qwen3 4B.
• For DeepSeek: a valid DeepSeek API key.
• For OpenAI-compatible mode: a compatible /models + Chat Completions endpoint; authentication is optional depending on the server.
• Git for diff tools, plus runtimes needed by project commands and MCP servers.
• One workspace root per extension session, selected at activation.

Native integration is verified on macOS. Windows and Linux still need native release checks before being claimed as verified platforms. Model output and command classification can be wrong; inspect changes and test results. Cancellation cannot undo completed process or MCP effects.

Development

Requires Node 24 and npm 11:

npm ci
npm run check
npm run package
npm run verify:package

Press F5 in VS Code to launch the development extension. Packaging writes dist/kova-local-<version>.vsix, using the name and version in packages/vscode/extension.package.json. The Marketplace extension identifier is owlcoder.kova-local; the current extension release is 0.3.1.

Local integration checks are smoke:ollama, smoke:vscode, smoke:vsix and smoke:ers.

Marketplace publishing

Publishing is intentionally manual. Build and verify the exact archive you intend to upload:

npm ci
npm run check
npm run package
npm run verify:package

Then upload dist/kova-local-0.3.1.vsix through the Visual Studio Marketplace publisher management portal for publisher owlcoder. No Marketplace PAT, repository secret or automatic publish workflow is required.

Setup · Providers · Architecture · Marketplace release · Release checks

MIT licensed.