Copilot Adapter Kit

Bring your own model to Copilot Chat — with a spend guard in front of it.
OpenAI, Anthropic, DeepSeek, Groq, OpenRouter, Ollama, LM Studio, vLLM, or your own endpoint.

▶ Watch the full tour (60s)  ·  Every screen  ·  Changelog

Why this exists

The first version of this extension did what the name says: it put any OpenAI-compatible model into the Copilot Chat picker. Then an overnight agent run on a configured provider burned roughly 100 million tokens, and nothing in the extension could say how — it counted nothing, capped nothing, and kept no record.

So 2.0 added the part that was missing. Every request now passes a guard that applies limits before it is sent, an audit log records what happened, and the status bar tells you where you are against the day's budget without opening anything.

The model picker still works the way it did. It just cannot run away any more.

Install

From the Marketplace: search Copilot Adapter Kit, or

code --install-extension salilvnair.copilot-adapter-kit

Quickstart

1. Click the ⬡ in the status bar → Settings
2. Providers → Add provider — pick a family, and the endpoint fills itself in
3. Paste your API key (it goes to the OS keychain, never to settings.json)
4. Models → Add model — the id your provider uses, e.g. deepseek-chat
5. Open Copilot Chat and pick it from the model dropdown

That is it. The defaults cap you at 2M tokens and $25 a day; change them in Spend Guard whenever they are wrong for you.

Spend Guard

Five limits, all applied before the request leaves, so a refusal costs nothing:

Every limit is typeable as well as steppable — 2M, 393k and $12.50 all parse. Set any of them to 0 to turn that one off.

The status bar carries the day: the mark alone when nothing has run, a percentage once something has, amber near the limit, red when it is refusing. The hover draws the day as a sparkline, because the shape says whether spend crept up all day or arrived in one burst at 3am — which a percentage cannot.

Turning protection off is a deliberate hold, not a click, and the status bar stays red the whole time it is off.

What it looks like

	Providers — every state at once: live, no key, key rejected, not running
	Models — grouped by provider; the switch is what the Copilot picker reads
	Audit Log — one row per model call and per panel action
	DB Explorer — the SQLite store underneath, a table at a time
	Ctrl K — providers, models, settings and actions in one search
	Git AI — branch, counts, per-file stats, and the cost before you send

All 27 screens →

What it does

Any OpenAI-compatible endpoint, plus Anthropic's native Messages API. Streaming, tool calling, agent mode, thinking blocks, rate-limit retry with backoff, and readable errors instead of raw 4xx.

Vision fallback — a model with no image support can still take images: the picture goes to a model that can see, and its description goes to the model you picked. Per-model or global.

Git AI — commit messages written by a model you chose, in a sidebar that shows the branch, the staged counts, the per-file stats and the cost of the request before it is sent.

An audit log on SQLite (via sql.js — no native compilation). One row per model call and per panel action, with the duration, the tokens and the cost. Prompt and response bodies are off by default, because that payload is your source code; turn them on in Developer Tools when you need them, and they are then recorded whole.

Developer Tools — Audit Log, Audit Config (which events get written down at all) and DB Explorer, for when "why did this behave differently to yesterday" needs an answer.

Security

• Keys live in VS Code SecretStorage. Never in settings.json, never shown back to the panel — the UI shows that a key exists, its health and its age.
• Nothing is ever sent anywhere to "test" a key. Reachability is probed with GET /models and no key attached; key validity is inferred only from real 4xx responses during actual use.
• The audit log never records a key or a request header — only the API path.
• Bodies are off by default. When on, they are clipped by nothing and kept for audit.retentionDays (30 by default).

Commands

Settings reference

Everything under copilot-adapter-kit.*. The panel writes all of it; this is here for the times you would rather edit JSON.

Providers

{
  "copilot-adapter-kit.providers": {
    "<uuid>": {
      "uuid": "<uuid>",
      "family": "openai",                      // which engine answers for it
      "name": "My OpenAI",
      "baseUrl": "https://api.openai.com/v1",
      "defaultApiPath": "/chat/completions",
      "modelApiPaths": { "codex-5.3": "/responses" },
      "modelAlias": { "gpt-4o": "gpt-4o-2024-08-06" },
      "visionFallback": "openai:gpt-5.2"
    }
  }
}

Providers are keyed by UUID and carry their family as a field — everything resolves them by family, so an entry without one can never be found.

Models

{
  "copilot-adapter-kit.models": {
    "<provider uuid>": [
      {
        "id": "deepseek-chat",
        "name": "DeepSeek Chat",
        "family": "deepseek",
        "maxIn": 128000,
        "maxOut": 8192,
        "image": false,
        "thinking": false,
        "toolCalling": 128,
        "pricing": "$0.27 / $1.10"
      }
    ]
  }
}

A map keyed by provider UUID, not a flat array.

Spend Guard

{
  "copilot-adapter-kit.budget.enforce": true,
  "copilot-adapter-kit.budget.dailyTokenLimit": 2000000,
  "copilot-adapter-kit.budget.dailyCostLimitUsd": 25,
  "copilot-adapter-kit.budget.maxInputTokensPerRequest": 200000,
  "copilot-adapter-kit.budget.maxOutputTokens": 0,       // 0 = the model's own maximum
  "copilot-adapter-kit.budget.maxTurnsPerConversation": 50
}

Everything else

Architecture

Every request from Copilot Chat passes through an interceptor chain, guard first:

Copilot  →  BudgetWarden  →  AuditRecorder  →  RateLimitGuard  →  ErrorWarden  →  DiagTracer  →  Engine
             refuse/allow     one row           429 retry ×3      readable       fingerprint     SSE

BudgetWarden runs first on purpose: a refusal has to happen before anything reaches the network, or it is not a limit, it is a report. AuditRecorder runs second so a refusal is recorded with its reason.

src/
  entry.ts              activation, commands, status bar
  kernel/               budget ledger, provider config, health, secrets
  mesh/                 engine contract, discovery, the interceptor pipeline
    engines/            openai (OpenAI-compatible), anthropic (Messages API)
  conduit/              the LanguageModelChatProvider VS Code sees
  crosscut/             budget-warden, audit-recorder, rate-limit, errors, tracing
  storage/              sql.js audit database
  panel/                webview hosts, status bar and menu
webview-ui/             React 19 + Vite + Tailwind 4 + @salilvnair/dui
  src/app/screens/      one file per destination
  src/app/devtools/     Audit Log, Audit Config, DB Explorer
  e2e/                  Playwright specs driving every screen
test/                   node suites: spend guard, packaging, contracts, db, health
scripts/                screenshots, demo video

Developing

npm install && npm --prefix webview-ui install
npm run compile          # webview build + esbuild bundle + tsc
npm run test:all         # node suites + Playwright

Press F5 for the Extension Development Host.

The extension host is bundled with esbuild into dist/extension.js — nothing is resolved from node_modules at runtime, which is what test/packaging.test.js exists to keep true.

Regenerating the media

npm run screenshots      # docs/screenshots/, every screen and tab
npm run showcase         # the demo video and GIF, into docs/demo/

Both drive the real UI against the Vite dev server with the fixture in webview-ui/src/app/fixture.ts, so every frame carries the amber sample-data bar. Nothing in any image or video is anyone's real usage. See scripts/demo-video/README.md for the recipes, the camera-move vocabulary and how to add a segment.

Troubleshooting

"SQLite did not start" — fixed in 2.0.1. Update; if it persists, Developer Tools names the error and everything else keeps working.

A model does not appear in the Copilot picker — it needs a provider with a key, and its picker switch on (Models). A model whose family has no provider behind it cannot resolve.

"Model does not support images" — set a vision fallback, globally or on that model, and reload the window.

Requests are being refused — that is the Spend Guard. The status menu says which limit and how far over you are; the Spend Guard screen is where you change it.

Something failed and the message is unhelpful — every command reports the command that failed with a Show log action, and the output channel has the stack.

Licence

MIT © salilvnair