Project Info

Corvus Agent

Devpost

Inspiration

AI coding agents are becoming capable enough to change real projects, but the trust model is still weak: a polished answer can hide incomplete work, unsafe commands, leaked secrets, or changes made outside the intended workspace. We built Corvus around a different question: what if an agent had to carry proof of what it did? Corvus is a local-first, proof-carrying AI agent workspace. It gives people the speed of an autonomous coding agent while keeping execution bounded, evidence visible, and final publication under human control.

What it does

Corvus turns a coding request into a supervised, inspectable run: Select a local project and an authenticated provider such as OpenAI Codex or Claude Code. Choose the model, thinking effort, and either Chat or Build mode. Review a server-authored safety preview that is cryptographically bound to the run configuration. Watch safe reasoning summaries, tool activity, and progress stream in real time. In Build mode, let the agent work inside a fresh isolated workspace rather than directly in the original checkout. Review the diff, test evidence, activity summary, detailed safety receipt, artifact hash, and secret-screening result. Download the screened project ZIP or explicitly prepare a branch and pull request for human review. Chat remains read-only. Build mode grants workspace-scoped write access while disabling user plugins, apps, and hooks by default. MCP servers are opt-in. Corvus never silently merges, force-pushes, or administers a repository. The interface adapts for Everyday or Developer users and Individual or Team workspaces without exposing every internal security primitive as configuration noise. The safety layer appears where it matters: before a run, while tools execute, when an action is blocked, and in the final receipt.

How we built it

A single authoritative Python core powers the CLI and authenticated FastAPI service. A React and TypeScript client provides the responsive web experience, while a Tauri and Rust shell packages the same interface as a desktop app. Owner-scoped Server-Sent Events provide reconnect-safe streaming. Every run binds the provider identity, model, project scope, authorization snapshot, sandbox policy, network policy, tool permissions, MCP consent, and idempotency inputs. Fixed-argument adapters launch supported local CLIs without shell interpolation. Build work happens in a managed scratch workspace; resulting files are bounded, screened for secrets, packaged with a SHA-256 manifest, and returned with audit evidence. OpenAI Codex was both our primary engineering collaborator and a first-class local runtime inside Corvus. We used it for architecture, implementation, security review remediation, CI repair, UX iteration, and end-to-end verification. The product then detects the user's installed Codex CLI, validates the selected model and effort, streams the run, and preserves the final evidence.

Challenges we ran into

The hardest problem was making safety real without making the product feel like a policy dashboard. We had to distinguish a genuinely failed run from an agent that completed its response but was waiting for confirmation, keep streaming responsive while hashing large workspaces, and make Windows workspace permissions compatible with a locked-down sandbox. Cross-platform packaging added another layer: the desktop app bundles a Python sidecar, React assets, and a Rust shell across Windows, macOS, and Linux. We also had to keep hosted-web claims honest: local mode hands off to the same-machine runtime, while cloud execution remains clearly labeled Preview until its E2B lifecycle is complete.

Accomplishments we're proud of

A working local Build path that edits and tests code inside an isolated workspace and returns a downloadable, screened artifact. Safety previews and terminal receipts authored by the backend rather than trusted from the UI. Streaming model output, safe work summaries, resumable events, cancellation, and retry behavior. Real local provider discovery, model selection, thinking controls, custom rules, MCP consent, skills, schedules, and project workflows. A responsive desktop/web experience with adaptive user and workspace profiles. A large automated certification suite across Python, web, Rust, Windows, macOS, Linux, Docker, secret scanning, and SAST. A release workflow that only publishes checksummed installers from a reviewed commit already on main.

What we learned

Security is most useful when it becomes understandable product feedback. “Protected Build,” a visible blocked-action explanation, and a final safety receipt communicate more than a long list of hidden controls. We also learned to treat completion as a verifiable state transition, not a confident sentence from a model.

What's next

Next we plan to add production E2B cloud sandbox lifecycle management, durable provider/autonomy/budget/kill-switch repositories, production Google OAuth and cross-device continuity, richer team collaboration, signed/notarized installers, and additional Build-capable provider adapters. Those items remain explicit roadmap work rather than being presented as finished features today. Submission snapshot The source reviewed for this Build Week entry is PR #16 on branch codex/demo-readiness-fix. It was approved and merged as commit 4d0244d and published as v0.2.0-build-week.1, with checksummed Windows, macOS, and Linux x64 artifacts. That release is the judging snapshot; later development on main is outside this entry.

Analysis

Compare with all teams

View

Metric

Figures cover GitHub contributors during the hackathon window. A co-authored commit counts in full for each author, so per-member totals add up to more than the whole-team figures.

Technology

Found in codeClaimed only
  • CSSIn code
  • FastAPIIn code
  • HTMLIn code
  • JavaScriptIn code
  • PythonIn code
  • ReactIn code
  • RustIn code
  • TypeScriptIn code
  • DockerClaimed
  • VercelClaimed

8 of 10 appear in the indexed code. 2 claimed on Devpost could not be matched to code, which may simply mean the tool leaves no trace in the repository.

AI coding agents

  • Claude CodeCommits

Detected from committed agent config files and commit authorship. Absence of a signal is not proof an agent was unused.

Codebase size

Source size

5.2 MB

Source files

398

Counts recognized source files only; vendored directories, binaries and lockfiles are excluded, so this is smaller than the repository on disk.

0 stars