ByokAgent — BYOK AI Coding Agent

Your keys, your models, your editor. Nothing in between.

ByokAgent is an agentic coding assistant for VS Code (and Cursor, via Open VSX) where the first screen is: paste a base URL, an API key, and a model ID. Any endpoint that speaks a known wire format works — no account, no gateway, no markup, no telemetry.

Why

Other open-source agent extensions support your own keys, but onboarding is built around their credit system and a long dropdown of hardcoded providers. If your provider isn't on the list, you're stuck. ByokAgent collapses the provider matrix into three wire formats:

If a provider isn't in the presets, fill the same four fields and it works.

Features (v0.1)

• Provider config — add/edit/delete providers: name, wire format, base URL, API key, model, context window, max output tokens, optional extra headers. One-click presets for OpenAI, Anthropic, Gemini, Nebius, Fireworks, Groq, Together, OpenRouter, Ollama, LM Studio.
• Model picker in the chat header; switch mid-conversation.
• Model auto-detect — "Fetch models" in the provider form pulls the real model list from {baseUrl}/models and shows it as a dropdown, so a wrong model slug can't slip through.
• Agent loop with read_file, write_file, apply_diff, list_dir, search (ripgrep), and run_command.
• Approval gate — every write shows a diff preview; every command shows the exact command. Approve / Reject / Always allow this pattern.
• Streaming chat in the sidebar, with cancel.
• Memory — conversations persist across reloads and tab switches, stored locally only.
• Cost meter — tokens in/out × per-model rates you enter.
• Secrets in the OS keychain (vscode.SecretStorage). Never in settings.json, never in the repo, never in logs.

Quick start

1. Install the extension, click the ByokAgent icon in the activity bar.
2. Pick a preset (or enter any OpenAI-compatible base URL) and paste your API key.
3. Type a task. Approve the writes and commands as they come.

Base URL:  https://api.studio.nebius.com/v1
API key:   your-key
Model:     moonshotai/Kimi-K3-Instruct

Local servers work too: http://localhost:11434/v1 for Ollama, http://localhost:1234/v1 for LM Studio — no key needed.

Security & privacy

• API keys are stored only in vscode.SecretStorage, backed by your OS keychain.
• Your code is sent only to the base URL you typed in — there is no other network endpoint this extension talks to. This is trivially verifiable from the source.
• The webview is sandboxed with a strict Content-Security-Policy: no remote scripts, no eval, nonce-only scripts.
• Anything that looks like an API key (sk-…, Bearer …, key=…) is redacted from error messages and tool output.
• Commands are denied by default. The always-allow list is per-workspace, and you can inspect or clear it with ByokAgent: Show Permissions.
• No telemetry. No analytics. No account.

Commands

Install in Cursor

ByokAgent publishes to Open VSX, which Cursor reads. Search for "ByokAgent" in Cursor's extension panel; if it hasn't propagated yet, install the .vsix from the GitHub release via Extensions: Install from VSIX.

Development

npm install
npm run build        # esbuild bundle → dist/extension.js
npm run typecheck
npm run test         # offline smoke tests of the adapters and diff engine
npm run package      # produces byokagent-<version>.vsix

Then F5 in VS Code to launch an Extension Development Host, or:

code --install-extension byokagent-0.1.0.vsix

Roadmap

• v0.2: git checkpoints + rollback, @file/@folder mentions, custom modes, per-project config file.
• v0.3: MCP client support, planning pass, prompt-caching headers, Ollama auto-discovery.

Non-goals: hosted gateway, credits, accounts, autocomplete/FIM, telemetry.

License

MIT