NoPilot

Provider-switching AI coding for VS Code. Use a persistent chat panel, inline completions, inline chat, commit messages, usage-aware provider controls, and local or remote model endpoints from one NoPilot menu.

NoPilot gives VS Code a lightweight AI coding workflow without locking you to one provider. Open the NoPilot Activity Bar entry, choose a provider first, then choose a model from that provider. You can use VS Code Language Models, Claude, OpenAI/GPT, Gemini, OpenAI-compatible APIs, and local or remote Ollama servers.

Highlights

• Activity Bar control center: Open NoPilot from the VS Code Activity Bar and jump to the chat panel, settings, provider/model selection, API keys, inline suggestion toggles, or commit message generation.
• Ask and Agent chat modes: Ask about the current editor, or run a bounded agent that reads workspace files, proposes reviewed changes, and requests approval for verification.
• Provider-first model picker: Choose the provider first, then pick from that provider's models instead of scanning one long mixed model list.
• Settings dashboard: Activate providers, set API keys, refresh live model lists, configure endpoints, and review request usage in one webview.
• Usage visibility: See current provider requests, total requests, top provider, and provider share in the status bar and settings dashboard.
• Inline completions: Get automatic ghost text with Fast, Balanced, or Rich quality profiles.
• Inline Chat: Select code, press Ctrl+I or Cmd+I, and ask NoPilot to edit or replace it in place.
• AI commit messages: Generate conventional or simple commit messages from your Git diff, or override the prompt with your own template.
• Remote and custom endpoints: Use OpenAI-compatible /v1 gateways and local, LAN, tunneled, or remote Ollama servers.
• Local-first secrets: Provider keys are stored through VS Code SecretStorage, not plain settings JSON.

Getting Started

1. Install NoPilot.
2. Open the NoPilot icon in the VS Code Activity Bar, or run NoPilot: Open Chat Panel or NoPilot: Open Settings from the Command Palette.
3. Configure a provider:
   • VS Code LM: Uses language models already available through VS Code.
   • Anthropic, OpenAI, Gemini: Save an API key, then choose a model.
   • OpenAI-Compatible: Set a baseUrl, save an API key, refresh models, then choose a model exposed by your server.
   • Ollama: Set a local or remote endpoint, refresh models, then choose a completion model.
4. Run NoPilot: Select Provider / Model from the sidebar, status bar, or Command Palette.
5. Start typing. Accept inline suggestions with Tab.

NoPilot Menu

The NoPilot Activity Bar entry opens a compact sidebar menu:

• Open Settings: Open the provider and extension settings dashboard.
• Open Chat Panel: Open the NoPilot chat panel beside the editor.
• Select Provider / Model: Pick a provider first, then a model from that provider.
• Set API Key: Save or change credentials through VS Code SecretStorage.
• Toggle Inline Suggestions: Enable or disable automatic ghost text.
• Generate Commit Message: Create a commit message from the current Git changes.

The status bar also shows the active provider/model with request usage and opens the same provider/model picker when clicked.

Provider And Model Selection

NoPilot separates provider selection from model selection:

1. Choose a provider such as VS Code LM, Anthropic, OpenAI, OpenAI-Compatible, Gemini, or Ollama.
2. NoPilot shows only models for that provider.
3. Pick a model, or open the full settings dashboard if setup is needed.

This keeps the model picker readable when several providers or live model catalogs are available.

Provider Setup

VS Code LM

Use this when you want NoPilot to call models already available through VS Code. You can leave nopilot.model empty to use the provider default, or choose a discovered model from NoPilot.

Anthropic, OpenAI, And Gemini

Use NoPilot: Set API Key or the settings dashboard to store credentials. Provider model settings are separate:

• nopilot.anthropic.model
• nopilot.openai.model
• nopilot.gemini.model

OpenAI-Compatible

Use this provider for custom OpenAI-style gateways, proxies, and self-hosted endpoints.

Example:

• nopilot.provider = openai-compatible
• nopilot.openaiCompatible.baseUrl = https://llm.example.com/v1
• Save your API key through NoPilot: Set API Key or the settings dashboard
• nopilot.openaiCompatible.model = gpt-oss:20b

NoPilot queries ${baseUrl}/models and sends chat-completions requests through the configured API root. Custom model names stay separate from official nopilot.openai.* settings.

Ollama

Set nopilot.ollama.endpoint to http://localhost:11434 or to a LAN, tunnel, proxy, or remote server URL. Refresh models from the settings dashboard, then choose a model with NoPilot: Select Provider / Model.

Remote Ollama mode can automatically use leaner automatic inline requests when latency or network behavior makes a server feel remote.

Keybindings

• Tab: Accept inline suggestions.
• Ctrl+I on Windows/Linux or Cmd+I on macOS: Open Inline Chat for the current selection.
• Source Control magic wand: Generate a commit message from the current Git diff.

Chat Panel

Open NoPilot: Open Chat Panel from the Activity Bar menu or Command Palette. The chat opens beside the editor.

Ask uses only the current file/selection, nearby editor text, and earlier Ask messages. It does not automatically include other open files or Agent tool results. Review Insert and Review Replace Selection target the URI, document version, and selection/cursor captured when the request was sent. Changing the active editor or selection does not retarget a response. Each edit opens an immutable diff, requires review and an explicit Apply Changes confirmation, and refuses documents changed since the request. Inline Chat uses the same review flow. NoPilot applies edits to editor buffers without explicitly saving them; VS Code Auto Save may save them. Undo is available.

Agent operates in one trusted local workspace folder (choose one in a multi-root workspace). It can list files, search literal text, read files, and propose exact replacements in up to five existing files. It must read a file before changing it. Each file diff is reviewed before a final approval applies the change set. Paths outside the folder, symlinks, hidden files, credential extensions, and dependency/build directories are rejected. Files are limited to 64 KiB; listing/search results are bounded and report truncation. New files, deletion, renames, remote workspaces, and arbitrary shell tools are outside this initial scope.

Verification can request only the root package.json scripts test, lint, compile, or build. A separate approval shows the actual script, working directory, buffers to save, and execution side effects. Only unchanged buffers previously approved in this run may be saved automatically; existing user edits must be saved separately. The script runs as npm --ignore-scripts run <name>, without npm pre/post hooks. Project scripts still execute with your user permissions and may change files or access the network: this is not a command sandbox. The exit status and up to 16,000 output characters go back to the model. Commands time out after 120 seconds. This initial command runner supports macOS/Linux.

All six providers use the existing completion interface with a validated JSON action protocol; native tool-calling support is not required. Models that cannot follow the protocol stop after bounded failures. A run has at most 12 model requests, plus limits on repeated actions, context size, and response waiting. Stop, closing the chat panel, or switching providers cancels the current run. Cancellation does not undo already approved edits, saves, or command effects. Provider-side cancellation is best effort, and a late response cannot trigger another action. Conversation and run state are in memory and do not resume after an extension restart.

For local validation, run npm test, npm run compile, npm run lint, and npm run build. npm run test:extension-host uses an installed VS Code (default macOS path; override VSCODE_EXECUTABLE) with a temporary profile and fixture workspace. Its model responses and approval choices are test doubles; actual VS Code document, diff, and workspace-edit APIs are exercised without provider calls. It does not load the user's installed extensions or settings.

Important Settings

You can customize NoPilot through VS Code Settings > Extensions > NoPilot or the NoPilot settings dashboard:

• nopilot.provider: Active provider for NoPilot requests.
• nopilot.model: Optional VS Code LM model override.
• nopilot.inline.enabled: Enable automatic inline suggestions.
• nopilot.inline.qualityProfile: Choose fast, balanced, or rich automatic inline behavior.
• nopilot.inline.pauseWhenCopilotActive: Pause automatic NoPilot suggestions when GitHub Copilot is active for the current language.
• nopilot.inline.debounceMs: Delay before requesting an automatic suggestion.
• nopilot.inline.maxPrefixLines: Maximum lines before the cursor to include as inline context.
• nopilot.inline.maxSuffixLines: Maximum lines after the cursor to include as inline context.
• nopilot.ollama.endpoint: Ollama server endpoint, local or remote.
• nopilot.ollama.remoteMode: auto, forced-on, or forced-off remote Ollama optimization.
• nopilot.ollama.model: Ollama completion model.
• nopilot.anthropic.model: Anthropic model.
• nopilot.openai.model: OpenAI model.
• nopilot.openaiCompatible.baseUrl: OpenAI-compatible API base URL.
• nopilot.openaiCompatible.model: OpenAI-compatible model.
• nopilot.gemini.model: Gemini model.
• nopilot.commitMessage.language: Commit message language, such as en, ko, or ja.
• nopilot.commitMessage.format: conventional or simple.
• nopilot.commitMessage.customPrompt: Optional custom prompt template. Supports {{diff}} and {{language}}, and overrides the preset format when set.

Example:

"nopilot.commitMessage.customPrompt": "Generate a commit message in {{language}} using the format '<type>(<scope>): <subject>'. Use one of feat, fix, docs, refactor, test, or chore. Include a Jira ticket from the branch name when present.\n\n{{diff}}"

Usage Visibility

NoPilot tracks request counts per provider in extension storage. The status bar and settings dashboard show:

• current provider request count
• total request count
• most-used provider
• provider share across configured providers

These counts are local convenience data for the extension UI.

Security And Data Handling

Your source code is sent only to the provider you select. Provider API keys are stored locally using VS Code SecretStorage. Ollama and OpenAI-compatible requests go only to the endpoint you configure.

• NoPilot sends prompts and code context only to the active provider.
• Provider API keys are stored locally in VS Code SecretStorage.
• Ollama requests are sent to the configured endpoint, which may be local or remote.
• OpenAI-compatible requests are sent to the configured baseUrl.
• This project currently does not implement custom telemetry or analytics collection.
• You are responsible for reviewing the data handling and retention policies of any third-party model provider you enable.

Do not commit real API keys, internal gateway URLs, or temporary tunnel URLs into workspace settings, screenshots, docs, or examples.

Troubleshooting

• Open NoPilot: Open Settings to confirm the active provider, API key state, endpoint, and selected model.
• Use NoPilot: Select Provider / Model to verify that the provider exposes selectable models.
• For Ollama, refresh models after changing nopilot.ollama.endpoint.
• For OpenAI-compatible servers, confirm that the configured baseUrl points to the API root, such as https://llm.example.com/v1.
• Check Output > NoPilot for provider state and request diagnostics.

Project Links

• Support: SUPPORT.md
• Contributing: CONTRIBUTING.md
• Code of Conduct: CODE_OF_CONDUCT.md
• Security: SECURITY.md

---

Built for developers who want AI coding help without giving up provider choice.