Vektor Chat — Ad-Funded AI (VS Code extension)

AI coding chat inside VS Code, paid for by sponsors instead of by you. While the model works on a reply, a large sponsor card fills the wait, then folds into a small sponsored line under the finished answer. No credits, no gates, no sign-up — the ads fund the chat.

• Zero configuration. Chat works right after install: the model API key lives on the Vektor server, never in your settings and never in this package.
• Agent mode. The model can read, search, create and edit files in your workspace, scoped to the workspace and respectful of unsaved editor changes.
• Free models, picked per turn. Auto routes each turn to the best free model for the job instead of pinning one.
• Ads are optional. Turn them off with vektor.adsEnabled; chat behaves exactly the same.

Run it (F5)

1. Open this folder in VS Code.
2. npm install (or bun install) once.
3. Press F5 — an Extension Development Host opens.
4. Command Palette → Vektor: Open AI Chat.

Model backend defaults to proxy — chat requests go to a Vektor chat server that holds the model API key server-side, so users never configure a key and there is nothing secret inside the extension to extract. The shipped vektor.proxyUrl / vektor.proxyToken defaults point at the Vektor Workers proxy (https://vektor-proxy.samton65.workers.dev), so chat works out of the box after install with zero configuration. The only other option in the model picker is mock, a no-key offline demo. Upstream providers are chosen on the server (see server/), never by the client.

Run your own proxy instead (see server/):

VEKTOR_UPSTREAM_API_KEY=<openrouter-key> node server/index.mjs

and override vektor.proxyUrl / vektor.proxyToken to point at it.

About vektor.proxyToken: it is a shared bootstrap token, not a secret. It ships in the package on purpose so the extension works out of the box, and the actual OpenRouter key never leaves the server. It is not a credential and cannot spend money on its own: every model served is free-tier (:free), so the provider's own free quota is what bounds usage, with the optional DAILY_REQ_CAP_PER_CLIENT / DAILY_REQ_CAP_GLOBAL allowances as a circuit breaker if you want one. Treat it as a public client identifier. If you run your own proxy, generate your own token, set it as VEKTOR_CLIENT_TOKEN, and ship that instead.

Free models it can route to: Auto (chosen per turn), MiniMax M3.1 Flash, North Mini Code, Dots 3 Note, Laguna S 2.1, Nemotron 3.5 Lightning.

The ad loop

 you send a message
                        │
                        ▼
      while the model works, the sponsor shows LARGE
      at the top of the streaming reply — the wait
      time is the ad slot (Freebuff-style)
                        │
                        ▼
            impression recorded once (deduped by impUrl)
                        │
                        ▼
        reply lands ──► the card folds into the small
        sponsored line under the answer (clicks tracked)

• Ads on (vektor.adsEnabled=true, default): every turn fetches a sponsor in parallel with the model call — the big card is the wait, the compact line stays under the finished reply. Requests are never blocked; there is no balance and nothing to run out.
• Ads off (vektor.adsEnabled=false): no ads are fetched or shown; chat works exactly the same (model traffic still flows through the Vektor proxy).
• Audit: impressions/clicks stream to the Vektor Ads output channel.

Settings

Ad server contract (when you set adServerUrl)

The client hits one normalized contract — the same one Freebuff's server mediates (POST /api/v1/ads, /ads/impression, /ads/click). Point it at your own waterfall (Gravity → ZeroClick → Carbon → first-party) and the extension switches from demo inventory to paid inventory with zero code changes.

• POST {base}/api/v1/ads body: { messages, sessionId, surface } where messages is text-only recent conversation (never file contents). Response: { ads: [{ adText, title, cta, url, clickUrl, impUrl }] }.
• POST {base}/api/v1/ads/impression body: { impUrl, surface } (a verified view; response body is ignored — the extension has no credit system).
• POST {base}/api/v1/ads/click body: { impUrl, surface }.

Roadmap

• [ ] Real agent tool loop (@codebuff/sdk agent runtime: terminal, diff apply) — workspace read/search plus file creation and editing (write_file / edit_file) already ship in the local tool loop
• [ ] Inline ads inside long responses (Freebuff pool: ≤4 unique ads per response, cycled for later slots)
• [ ] Composer-adjacent rotating slot (60s rotation, idle pause)
• [x] Session caps to bound inference spend — now enforced server-side by the proxy's daily allowances (DAILY_REQ_CAP_PER_CLIENT / DAILY_REQ_CAP_GLOBAL), so no credit system needs to live in the client
• [x] Package (.vsix) for the VS Code Marketplace
• [ ] Publish to Open VSX and a self-hosted side-load

Privacy stance

Targeting context sent to an ad server is conversation text only — never file contents, never repository data. Sponsors don't see your code. (Same line Freebuff draws in its privacy policy.)