Inspiration
Coding agents are fast, but they write to your disk, run code on your machine and claim "done" without proof. We wanted an agent we would trust on a real repository: every change reviewed, every test isolated, every claim backed by evidence.
What it does
Polaris is a local-first coding agent written in Rust.
- One durable session, four frontends. A kernel keeps each session as an ordered event log. A terminal UI, a desktop app (Slint), a web UI (Leptos) and a daemon all show the same session, and it replays after a restart.
- Nemotron proposes, you decide. Replies come from NVIDIA Nemotron chosen by effort level. Every file change waits as a diff for Apply or Decline;
/diffand/undokeep the history. - System One. A local GLiNER2.5 classifier (ONNX, CPU) screens each change for test tampering, leaked secrets and intent mismatch. A gate decides alone only when calibrated at ≥ 99 % precision; otherwise it asks you.
- NVIDIA OpenShell sandbox. The model can ask to run tests; you approve the exact commands; they run in an isolated OpenShell sandbox without network, and Proof of Done blocks a false "done".
- Local toolkit.
/nv-profileand/nv-verify(CUDA advisor),/hack(races, blocking calls in async code, secrets), local and web search, CI in Docker, a research loop. - Built for Nebius. Token Factory (chat by effort, embeddings, fine-tuning), Contree SWE sandboxes, Data Lab and AI Cloud are integrated behind one settings file.
How we built it
95 Rust crates in one workspace (Rust 1.99): an event-sourced kernel, capability-scoped tools, one OpenAI-compatible client for Token Factory, OpenRouter and local servers, ONNX Runtime for System One, NVIDIA OpenShell 0.1.4 on Docker in WSL2 for the sandbox, and Ratatui, Slint and Leptos for the frontends. 2,630 tests passed in the last full run, and seven performance budgets are measured.
Challenges we ran into
- Nebius access. Our team is in Ukraine, where Nebius services were not available to us, so we could not run Token Factory, Contree or AI Cloud live. We ran Nemotron through OpenRouter (and Nemotron-Mini locally) and wrote the Nebius integrations against the published APIs: ready for a key, not yet run live.
- Whole files vs. a 512-position encoder. Our first System One calibration measured almost nothing: full before/after files exhausted the worker's memory. We switched to bounded changed-line inputs.
- OpenShell on Windows: WSL networking, broken IPv6 for apt, and a gateway that binds only seconds after the VM boots.
Accomplishments that we're proud of
- Live OpenShell acceptance 4/4: real tests pass with Proof of Done, while zero tests, a failing command and too few tests are all blocked, and a replayed approval is refused.
- System One calibrated with the live worker: 100 % precision on two of three safety gates (400-case synthetic set).
- Nothing in Polaris writes to your workspace without a reviewed diff.
What we learned
Agent safety is mostly plumbing: durable logs, exact review revisions, isolation and honest calibration matter more than prompts. Small models make excellent gatekeepers once you measure them.
What's next for Polaris
Run Token Factory, Contree and Data Lab live, calibrate System One on real changes, and fine-tune Nemotron on proved trajectories.
Log in or sign up for Devpost to join the conversation.