Flint — Foundry Local Interface (FLInt)

Flint (also styled FLInt) is a lightweight, desktop GUI for Microsoft Foundry Local.

It provides an intuitive interface for:

• Model catalog browsing/search/filter (STT-only for audio via metadata), hardware-aware recommendations, download/load/unload + progress, and WinML/accelerator preference support
• Chat (streaming) with conversation sidebar, persona selector, localStorage persisted history, system prompt controls, stop/cancel, and basic vision image support (emerging in SDK)
• Audio transcription (mic + file) using STT models with copy/download transcript actions
• Diagnostics with service controls, endpoint visibility/copy snippets, and active execution provider inventory
• Learn section + local endpoint
• Bundled runtime (sidecar eval for clean prod builds), first-run guidance

Everything runs locally on your device by default.

Status

Flint is now at an MVP 0.1-ready baseline for dogfood/public-preview use:

• Models: catalog/search/filter + download/load/unload + model metadata details
• Chat: streaming + conversation UX + thinking-trace rendering + cancellation path + persona presets
• Audio: STT model selection + mic/file transcription + transcript copy/download
• Diagnostics: service controls + endpoint/snippet visibility + execution provider inventory
• Test baseline: unit + sidecar protocol tests + coverage gating

See release evaluation + roadmap: RELEASE_ROADMAP.md

MVP 0.1 Interface

Main window

Model selection

Chat window

Known Limitations (0.1)

• Multi-endpoint orchestration is not implemented; most flows assume one active endpoint/model lane.
• Diagnostics export and cache operations are functional but not yet fully hardened/polished.
• Security posture is improved but not yet least-privilege complete (shell capability scope will be tightened in 0.2).
• Test coverage is foundational (good baseline), not full UI/E2E depth yet.
• Some audio features may not work as desired.

Working Features (MVP 0.1)

Models

• Browse and search local model catalog
• Metadata-driven model filtering (including STT-focused audio lane)
• Download/load/unload models with progress feedback
• Model details panel (task/context/capabilities)

Chat

• Streaming assistant responses
• Conversation sidebar with persisted history
• Persona selector and editable system prompt
• "Thinking" trace rendering in chat responses
• Stop/cancel active generation
• Basic image attach flow for vision-capable models

Audio

• Microphone transcription
• File transcription
• STT model selection
• Copy transcription text
• Download transcription text

Diagnostics

• Sidecar/service status and controls
• Local endpoint display
• Copy-ready endpoint snippets
• Execution provider list and active provider visibility

Testing & quality baseline

• npm run check static/type validation
• Vitest unit tests for core utilities
• Sidecar protocol smoke test
• Coverage thresholds in CI/dev test runs

Tech

• Tauri 2 (Rust + WebView)
• Svelte 5 + SvelteKit + TypeScript
• foundry-local-sdk (primary) + foundry CLI (supporting)

Getting Started (Development)

npm install
npm run tauri dev

Prerequisites

• Node.js + npm
• Rust + Cargo (for Tauri)
• The runtime is bundled in builds (no separate install needed for end users)

On Windows prefer the WinML variant for best acceleration:

npm run setup:winml

Building & Packaging

npm run tauri:build
node scripts/verify-bundle.js

Running the built app

• During development: npm run tauri dev (recommended, hot reload, uses sidecar)
• Test a production build without installing the MSI: scripts\run-built.bat (or pwsh scripts/run-built.ps1, or double-click the .exe in src-tauri/target/release/)
• Distribute to users: Use the generated MSI (src-tauri/target/release/bundle/msi/Flint_*.msi) or NSIS installer. These create proper Start Menu entries, uninstaller, etc.

See .github/workflows/ for CI (PR checks + builds) and Release (tagged builds create GitHub releases with artifacts for Windows/macOS).

Icon asset generation

To refresh app icon assets from a source image:

pwsh .\scripts\flint-icon-generator.ps1 -SourceImage .\static\flint-master-1024.png -RepositoryRoot .

This updates icon assets in static\ and src-tauri\icons\ (including .ico). SVG generation is intentionally excluded because it was producing raster-wrapped output rather than true vector artwork.

Sidecar evaluation (for native SDK issues): Direct use of foundry-local-sdk (JS + native) from the Svelte frontend has bundling challenges:

• Vite externalizes Node core modules (fs, url, module, etc.) during tauri build.
• Native prebuilts + core DLLs are copied via resources, but resolution can fail in the final bundle.
• See sidecar/foundry-sidecar.js for a prototype stdio-based sidecar wrapper (JSON protocol). Use @tauri-apps/plugin-shell to spawn and communicate from frontend.

Recommendation: For production, evaluate moving logic to a Rust sidecar (or directly in src-tauri using the Rust foundry-local-sdk crate). This gives clean native bundling and no JS FFI headaches. Current JS approach is kept for rapid iteration per original design spec.

Sidecar readiness: The sidecar now emits { "ready": true } immediately after listener setup and lazy-loads the SDK only on the first init command. resolveResource + cwd + NODE_PATH are used for robust script/SDK resolution in dev + prod bundles. Stderr is captured. Timeout is 10s.

Note on distribution: Because the sidecar runs node <script>, end users of built installers still need Node.js installed and on PATH. (A self-contained .exe sidecar via pkg or a Rust port would remove this.)

Run npm run tauri:build (may require full MSVC/Rust env) and node scripts/verify-bundle.js.

Vision / Multimodal

• Added basic support in Chat: image attach (base64) when model alias suggests vision/multimodal.
• Messages use OpenAI-style content array for images.
• Helper getVisionModels() in SDK (based on task/capabilities).
• Full support is emerging in the local SDKs (chat models with vision capabilities exist in catalog like certain Phi/Qwen variants; image input via the compatible API). Currently best-effort; test with actual vision models. No dedicated vision tab yet.

CI/CD

• PRs: Lint, checks, cross-platform debug builds
• Releases: Push vX.Y.Z tag → automated release with installers

Design

• FLINT_DESIGN_SPEC.md
• IMPLEMENTATION_PLAN.md
• MVP_FEATURE_COMPLETION_PLAN.md
• RELEASE_ROADMAP.md

License

MIT

Contributing

See the design spec for scope and roadmap.

---

Built to make Foundry Local approachable while staying true to its local-first roots.