A native desktop hub for Git workspaces, repository state, BYOK AI providers, and Cognee-backed memory.
Memory powered by
Cognee
Overcode is a native desktop application that consolidates local Git state, GitHub, and GitLab into a single operator workspace. It is built for engineers who maintain multiple repositories, monitor pull requests and merge requests, and switch between branches, worktrees, and stashes during a normal working day.
The data plane is local. Repositories are read directly from disk through simple-git in an isolated worker process. Remote provider data is pulled per-account through user-scoped OAuth and cached locally. AI calls are explicit, routed through the selected provider, and recorded in a local audit log. Cognee is used as an optional repository memory layer for approved summaries and structured facts.
There is no Overcode backend.
Overcode is entered in the Developer Tools category. The application existed before the July 13, 2026 submission window; the judged Build Week contribution is a focused GPT-5.6-assisted extension rather than a claim that the entire product was built during the event.
- Replaced a shallow, duplicated memory workflow with one tested repository-memory lifecycle for recall, fact collection, stable deduplication, remember, improve, and forget behavior.
- Migrated repository briefing, impact analysis, commit assistance, issue triage, standup, worktree comparison, pull-request, repository, dashboard, and settings surfaces to that shared boundary.
- Added repository-memory and theme-input tests.
- Removed mandatory route, keyboard-command, AI-panel, and list entrance delays; preserved causal pointer feedback; and made reduced-motion behavior informative without decorative loops.
- Narrowed the worktree-comparison ARIA live region during independent acceptance review so it does not contain an interactive control.
The implementation was completed in Codex with gpt-5.6-sol at xhigh reasoning in session 019f7679-b2ee-7961-bdac-6ec8a426ca31. GPT-5.6 analyzed the existing architecture, implemented the consolidated module and caller migration, refined the interaction motion, and added tests. The human selected the architectural boundary, constrained memory to approved repository summaries and structured facts, chose immediate keyboard behavior with spatial feedback only for pointer input, and made the final accessibility correction. The same GPT-5.6 session reviewed that correction, and /feedback was submitted successfully.
Acceptance verification for this contribution:
| Check | Result |
|---|---|
| Unit tests | 20 files and 145 tests passed |
| TypeScript | npx tsc --noEmit passed |
| Lint | Passed with zero warnings |
| Production package | npm run build produced the Linux AppImage |
| Patch hygiene | git diff --check passed |
The Overcode website distributes the prebuilt Linux AppImage from the public GitHub releases. The Build Week contribution is packaged as v0.1.2, preserving the historical v0.1.1 artifact from before the event. Local repository and worktree views need no account, while AI providers, Cognee, GitHub, and GitLab are optional integrations.
| Capability | What it does |
|---|---|
| Local workspace scan | Recursively discovers Git repositories beneath operator-configured roots. |
| Repository dashboard | Shows branches, uncommitted files, stashes, worktrees, recent commits, and file-system refresh state. |
| Activity feed | Unifies commits, pull-request transitions, issue updates, and pipeline outcomes across linked accounts. |
| Pull-request console | Combines GitHub PRs and GitLab MRs with hunk-level navigation. |
| Issue console | Shows cross-provider issues with assignee and linked-PR context. |
| AI panel | Runs impact analysis, commit assistance, repository briefings, stash annotation, worktree comparison, code explanation, issue triage, and daily standup generation. |
| Cognee memory | Remembers approved repository summaries, risk notes, module facts, decisions, and recall context across sessions. |
Each AI feature routes through electron/lib/ai-runtime.ts and validates structured responses before rendering them. On parse failure the response is repaired once; if repair fails, the UI falls back to a local-data-only envelope instead of surfacing a raw exception.
Three isolated process tiers:
- Main process (electron/) - IPC handlers, OAuth callback servers, AI provider adapters, Cognee adapter, and
electron-storepersistence at~/.overcode/config.json. - Preload bridge (electron/preload.ts) - typed
window.apiexposed throughcontextBridge. - Renderer (src/) - React 18 + TypeScript, Zustand state, and vanilla CSS design tokens.
simple-git and chokidar run inside an Electron utilityProcess worker so the main process and renderer remain responsive against large repositories.
Architecture diagrams are documented in ARCHITECTURE_DIAGRAMS.md and the accompanying Excalidraw source files.
Overcode reads local files and watches them continuously. A hosted web app would need per-folder browser permission, has weaker file-watch behavior, and would require a backend for OAuth callbacks and token storage.
| Concern | Native desktop behavior |
|---|---|
| Local repository access | Direct OS-gated filesystem access. |
| File-system change events | Continuous chokidar watcher in a utility process. |
| OAuth callback | Local callback server on 127.0.0.1, started on demand. |
| Credential storage | electron-store on the operator's endpoint, encrypted through Electron safeStorage when available. |
| Offline behavior | Local Git views remain usable without network. |
| Deployment | One desktop artifact per platform. |
| Component | Minimum |
|---|---|
| Operating system | macOS 11, Windows 10 22H2, or glibc 2.31+ Linux |
| Memory | 4 GB available |
| Disk | 500 MB for the application, plus repository metadata cache |
| Network | Outbound HTTPS to OpenRouter, OpenAI, Anthropic, Gemini / AI Studio, Cognee if configured, GitHub, GitLab, and any configured enterprise instances |
| Accounts | Optional AI provider API keys, optional Cognee endpoint, optional GitHub and GitLab OAuth applications |
git clone https://github.com/Timidan/overcode.git
cd overcode
npm install
cp .env.example .env
npm run devFor a clean development run:
./scripts/clean-start.shTo remove all stored state, delete ~/.overcode/ or run:
./scripts/clean-start.sh --reset-user-dataOvercode uses bring-your-own-key AI providers. You can save keys for OpenRouter, OpenAI, Anthropic, and Gemini, then choose one active provider and model under Settings -> AI Providers.
OpenRouter is the default because it exposes a broad model catalog, including free and paid models. Direct provider keys are supported for users who prefer billing through OpenAI, Anthropic, or Google AI Studio.
Paid models are allowed. Overcode marks provider-billed models before activation and does not charge for model usage.
Stored Settings values take precedence over environment variables.
- Open Settings -> AI providers.
- Choose OpenRouter, OpenAI, Anthropic, or Gemini.
- Paste the provider API key, then click Save credentials.
- Choose a catalog model or enter a manual model ID.
- Acknowledge provider billing if the selected model is paid or pricing is unknown, then click Activate provider.
Credentials are persisted to ~/.overcode/config.json and encrypted with Electron safeStorage whenever the operating-system keystore is available. On systems without a keystore, values are stored as plaintext in that file.
| Provider | Environment variables |
|---|---|
| OpenRouter | OPENROUTER_API_KEY, accepted alias OPENROUTER, optional OPENROUTER_MODEL, optional OPENROUTER_BASE_URL |
| OpenAI | OPENAI_API_KEY |
| Anthropic | ANTHROPIC_API_KEY |
| Gemini / AI Studio | GEMINI_API_KEY, accepted alias GOOGLE_API_KEY |
Run the generic smoke path with:
node scripts/smoke-ai-provider.mjsRun the OpenRouter-specific check directly with:
node scripts/smoke-openrouter.mjsCognee memory is enabled when an endpoint and API key are present. COGNEE_API_URL is the canonical endpoint variable, while COGNEE_SERVICE_URL and COGNEE_BASE_URL are accepted aliases.
| Key | Required | Notes |
|---|---|---|
COGNEE_API_URL | one endpoint key | Preferred Cognee API base URL, without a trailing route. |
COGNEE_SERVICE_URL | one endpoint key | Accepted alias. |
COGNEE_BASE_URL | one endpoint key | Accepted alias. |
COGNEE_API_KEY | yes | Sent by the main process; never exposed to the renderer. |
Cognee runs in context-only mode. It stores approved repository memory and recall context, not raw source, raw diffs, prompt bodies, credentials, OAuth tokens, or .env values.
If you want the unified PR, issue, and pipeline console, register OAuth applications and enter the client credentials under Settings -> Integrations.
| Provider | OAuth callback URL | Where to register |
|---|---|---|
| GitHub | http://127.0.0.1:3000/callback | GitHub developer settings |
| GitLab | http://127.0.0.1:3001/callback | GitLab profile applications |
Both providers are optional. Local Git views work without either.
npm run buildUseful platform-specific commands:
npx electron-builder --linux AppImage
npx electron-builder --win nsis
npx electron-builder --mac dmgOutput is written to release/0.1.2/.
| Boundary | Implementation |
|---|---|
| Secrets in version control | .env is ignored. Keys, tokens, and OAuth client secrets are not committed. |
| Credential storage | OAuth and AI provider credentials are written by the main process. The renderer cannot read the raw credential store or secret blobs through IPC. |
| OAuth flow | A random state parameter is validated on callback. Local callback servers bind to 127.0.0.1 and are started only for a single authorization round-trip. |
| Renderer trust boundary | Token material, raw OAuth responses, and raw AI request bodies do not cross the preload bridge. |
| External URL handling | shell:open validates URLs against an allowlist before calling shell.openExternal. |
| Repository data | Repository content stays on the operator's endpoint except for explicit AI calls and approved Cognee memory writes. Audit metadata is stored locally without prompt or response bodies. |
- Consolidated the repository-memory lifecycle and migrated every AI and repository caller to the shared boundary.
- Added repository-memory and theme-input coverage; 145 tests pass in the release candidate.
- Removed mandatory entrance delays from keyboard, route, AI-panel, and list interactions.
- Refined reduced-motion feedback and corrected the worktree-comparison ARIA live region.
- Desktop Git workspace hub with GitHub and GitLab OAuth.
- BYOK-provider-backed AI panel and inline PR analysis workflows.
- Cognee-backed repository memory for remember, recall, improve, and forget flows.
- Linux AppImage, Windows NSIS, and macOS DMG build targets.
Overcode is available under the MIT License.