BuilderForce for VS Code

A codebase-aware BuilderForce AI coding agent in your editor sidebar.

Status: working agent. Sign in, point it at your repo, and ship — it edits code with your approval and runs your team's work without leaving the editor.

Features

• Ship code by chatting — describe what you want and the agent finds the right code, edits it (sandboxed to the open folder), runs your tests/build/lint to verify it works, and can commit, push, and open a PR — going from idea to shipped without leaving the editor. Every change and command asks for your approval first.
• Run your team's work from the editor — open, update, and dispatch projects, tasks, and OKRs directly in chat, and kick off a run on a ticket and follow its status — no dashboard. A Work Inbox in the sidebar shows what needs you (pending approvals) with one-click starts to review your pull requests and their CI, fix a production error from the error straight to the code, or open a pull request for your changes.
• It knows your codebase — scans the open folder once (cached) and grounds every answer in your project's real structure, so suggestions fit your code. Refresh with Rescan Codebase.
• Long chats keep going — busy conversations that pull in big results (like a full task list) used to fill up and stall the assistant part-way through. Now the chat keeps only what matters as it works, so it stays responsive and finishes the job instead of dying after a few steps.
• Spends less for the same work — a small request runs as one quick stage on a cheaper model; a big one gets the full plan → code → review pipeline, each stage on its own model. Questions such as "where is X defined" are answered from the symbol index with no model call. The agent can hand code reading to a model on your machine (local_research), starts from the files your past runs mapped for that area, and keeps its prompt stable so provider caching applies. Set a per-run spend cap, and open Pipeline route report to see what each route cost over the last 30 days.
• Attach files and context that the model actually reads — attached text files, Add context and #file references are put into the message within a size budget. Long prose (Markdown, logs, text) is compressed first; code never is. Binaries upload to your account.
• Show it a screenshot — paste or attach an image of a bug, error, or design and the agent acts on what it sees.
• See exactly what it did — and why it stopped — every step (its reasoning, each tool call with input/output, file edits, and errors) is visible, so you can trust and verify the work. One click copies the whole transcript, now with a plain-English Diagnostics summary: if a chat ever stalls, the copy tells you the likely reason at a glance — perfect to drop into a bug report or hand to support. Another click runs connection diagnostics.
• Pick up anywhere — your conversations are the same ones in the BuilderForce web app, so you can start in the editor and continue in the browser (or the reverse).
• Secure sign-in — one-click browser login (or paste-key); your key lives in the OS keychain, never in settings or logs.
• Your gateway, your models — defaults to https://api.builderforce.ai; override builderforce.baseUrl for self-hosted, and pick any model from the live pool or let the gateway choose.
• Or run the model on your own machine — turn on builderforce.localModels.enabled and everything your Ollama install has pulled, plus whatever your FreeToken engine is serving, appears in the model picker under On this device. Pick one and the turn runs on your hardware with the same tools, edits, and approvals as any other chat: nothing is spent from your plan, and it keeps working with the network down. See Running the model on your own machine for what does and doesn't stay local.

Getting started

1. Install the extension and open the folder you want to work in.
2. Click the BuilderForce icon in the Activity Bar.
3. Click Sign In (a browser opens to approve, or paste an API key from https://builderforce.ai).
4. Ask it to build or change something — it will scan the repo, propose edits, and apply them on your approval.

Running the model on your own machine

The extension speaks the OpenAI-compatible API that both runtimes expose, so no adapter or extra tooling is involved.

1. Start a runtime and leave it serving:
   • Ollama — ollama serve, with at least one model pulled (ollama pull qwen3:8b). Listens on http://127.0.0.1:11434. Any Ollama model works, including a JetBrains Mellum 2.1 GGUF — ollama pull hf.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF:Q4_K_M (~8.1 GB) — which shows up as its own row under On this device.
   • FreeToken — start the engine for a model in the desktop app. Listens on http://127.0.0.1:1919.
2. Turn it on — tick BuilderForce › Local Models: Enabled in settings. Only change the two URL settings if you moved a runtime off its default address; both the bare address and the /v1 form work.
3. Pick a model — run BuilderForce: Pick Model and choose a row under On this device. Each row is labelled with the runtime serving it.
4. Chat as usual. To go back to the gateway, pick any other model.

A runtime that isn't running simply contributes no rows — the picker still opens, so you can run only one of the two.

What stays on your machine, and what doesn't

The inference is entirely local. Your prompt, the code context gathered from your repo, and the model's reply are exchanged only with the runtime on your machine. They are not sent to the BuilderForce gateway or to any model provider, and the turn costs nothing from your plan.

Your transcript is still a platform record. The sidebar chat saves its conversation to your BuilderForce account exactly as it always has — that is what makes it the same conversation in the web app — so the text of a local chat is stored server-side like any other. The model that produced the text has no bearing on where the text is kept. The @builderforce participant and the codebase scan store nothing when you are signed out.

If you need a turn that touches no server at all, use @builderforce in the VS Code chat view with a local model picked and no account signed in.

Settings

Development

npm install
npm run compile      # or: npm run watch
# Press F5 in VS Code to launch the Extension Development Host

Publishing

See PUBLISHING.md for the dual-registry (VS Code Marketplace + Open VSX) release process.

License

MIT © Sean Hogg / BuilderForce