Introducing uppcoder agent: AI coding that lives in your terminal
Most AI coding tools today want one of two things: to live inside an editor extension that ships you Electron and 400MB of dependencies, or to wrap you in a web app where your code and prompts pass through someone else’s servers by default. We wanted neither — so we built uppcoder agent.
It’s a single, native Rust binary. It runs in your terminal. It talks to whatever model you point it at. And it doesn’t phone home.
What it actually is
uppcoder agent is a full-screen TUI coding assistant. You type a request, it streams a response, and it can execute tools on your filesystem — read, write, edit, glob, grep, run shell commands, fetch the web, manage tasks — with a per-tool permission system so nothing happens without your say-so.
You can drop into classic line mode (--classic) for scripts or simpler terminals. Both modes share the same engine.
The differentiators
Bring your own model. No vendor lock-in. Configure any of:
- Anthropic Claude (the default for most users)
- Ollama — direct or via OpenWebUI, for fully local inference on your own machine
- OpenRouter — one key, every model
- Any OpenAI-compatible endpoint — including self-hosted vLLM or Together
Switch providers per-session with --model, or set defaults in config.toml. We don’t take a cut, and we don’t track which model you use.
It runs where you work. Editor extensions break your flow when the model thinks for 30 seconds; the spinner is your interface. In a terminal next to your shell and your code, the model thinking is just one more pane to glance at while you keep working.
Permissions you can audit. Every tool call asks before it runs — approve once, approve for the session, or deny. The default is paranoid. No silent rm -rf from a confused model.
MCP-native. Plug any Model Context Protocol server in via config and the agent gets new tools dynamically. Filesystem, GitHub, Slack, your own internal APIs — anything with an MCP wrapper just works.
Token usage tracked locally. A bundled SQLite database keeps per-session, daily, weekly, and monthly stats. You can see exactly what you’ve spent, broken down by provider and model.
Persistent chat history. Conversations survive restarts. So does your active task list — the agent’s todo panel is shown in a side panel and tracked across sessions.
Native Rust. Cold start is instant. Memory footprint is measured in megabytes, not gigabytes. No Node runtime, no Electron, no virtual environments.
What it isn’t
Right now, uppcoder agent is not open source. We’re shipping it as a commercial product with an early-access invite. If you want to try it, get in touch — we’ll send you a build and onboard you directly.
Some features are still rough around the edges (the keypress timing on permission dialogs is something we’re actively polishing). You should expect to be talking to us as we iterate. That’s part of the point.
Who it’s for
- Engineers who already live in a terminal and want AI assistance there, not in another window.
- Teams that need to keep code and prompts on their own machines (or their own LAN) for compliance or sanity reasons — Ollama support means fully local operation.
- Anyone tired of paying a token markup to a wrapper company.
Try it
Get in touch for early access. Tell us what stack you work in and what model you’d like to start with, and we’ll get you running the same day.