Skip the AI 'Harness': Manual Context for 15+ Chatbots

What it does

CodeWebChat is a VS Code extension that generates XML-formatted prompts from your hand-picked file selection and sends them to over 15 chatbots, either via direct API calls using your own key or through a browser extension that autofills the web UI. It also offers inline code at the cursor, commit-message generation, voice input, and workspace checkpoints, all while claiming zero telemetry.

The interesting bit

The project flips the agent model: instead of the AI pestering a tool harness to read files iteratively, you curate the context directly in the native file explorer and the model receives everything upfront. The README notes that this design makes fresh chats the natural workflow, and ordering files by recent edit history helps exploit prompt caching across tasks.

Key highlights

• User-managed context from the native file explorer; no tool-calling round-trips required.
• XML-formatted prompts for edits, questions, cursor insertion, and natural-language file discovery.
• Supports API calls via BYOK or self-hosted providers like Ollama and LM Studio, plus a browser extension that autofills web chatbots over a local websocket.
• Zero telemetry; requests go directly to configured model providers.
• Checkpoints, intelligent update integration, and voice input round out the feature set.

Caveats

• The “Apply response” button beneath chatbot replies is only a copy-to-clipboard alias, not automatic code extraction.
• Remote SSH development requires forwarding port 55155 for the browser-autofill websocket.
• Autofill depends on installing a separate browser extension; without it, you are back to copy-and-paste.

Verdict

Developers who distrust black-box agent loops and would rather curate their own context will appreciate the transparency. If you are looking for a fully autonomous coding agent that reads its own files, this is explicitly not it.