VICODE + CHITOS: Run a Full Coding Agent Without Your Code Leaving the Building
Run a full coding agent without your code leaving the building
VICODE is an on-prem coding agent: a self-hostable, Claude-Code-style harness whose default backend is an open model - Kimi-K3, served through the Hugging Face router. The design goal is not "another AI IDE." It is data sovereignty. Your source, your prompts, and the agent's intermediate reasoning stay on infrastructure you control.
Why on-prem is the whole point
Most coding assistants are thin clients in front of a closed API. That is acceptable right up until your source code is the asset you cannot leak - banks, hospitals, defense, public-sector, anyone under strict data-residency rules. For those teams, "upload the repository to a third-party endpoint" is a non-starter.
VICODE flips the default:
| Typical hosted agent | VICODE | |
|---|---|---|
| Backend model | Proprietary API | Open model (Kimi-K3 default) |
| Where your code goes | Vendor cloud | Your own infra |
| Self-hostable | No | Yes |
| Model swappable | Rarely | Yes |
Because the default backend is an open model, you can run the whole loop - edit, run, read errors, iterate - inside your network with no mandatory egress. Prefer a different open model? The backend is swappable.
One engineering note
Hosted reasoning models can burn their entire token budget on hidden chain-of-thought and return an empty completion. We hit exactly this and fixed it by controlling the reasoning budget (a low reasoning-effort setting) so the model spends tokens on actual code output. If you self-host an agent around a reasoning model, budget the thinking explicitly - it is an easy way to lose all your useful tokens.
CHITOS / MOUSE: the red-team half
Shipping code you did not fully write raises an obvious question: is it safe? CHITOS (also called MOUSE) is our AI security scanner and red-team agent - it probes applications and code for weaknesses rather than just generating them. There is a public demo at mouse.1street.ai so you can see the behavior before trusting it on anything real.
Together the two form a loop: VICODE writes and edits under your control, CHITOS pushes back on what was produced.
Honest scope
VICODE is a coding harness, not a magic autonomous engineer - it still needs review, tests, and a human in the loop, and open-model backends vary in capability. "No data egress" describes the on-prem deployment you run; if you point the backend at a hosted router endpoint instead, data goes wherever that endpoint lives - deploy accordingly. CHITOS/MOUSE is offered as a public demo (demo-stage): it is an assistant for finding issues, not a certification, a compliance sign-off, or a guarantee that code is secure. Treat both as tools that raise your baseline, not as replacements for security engineers or code review.
More: https://vidraft.net
Top comments (0)