Flint

The desktop control plane for Microsoft Foundry Local.

Manage models on your machine, chat and transcribe locally, pit models against each other in the Model Arena, and expose an OpenAI-compatible endpoint to the coding tools you already use — without sending prompts to a cloud by default.

Also styled FLInt (Foundry Local INTerface). Product name: Flint.

---

Why Flint?

You want…	Flint gives you…
Privacy	Inference stays on-device. Network bind defaults to loopback; non-loopback requires an explicit choice and confirmation.
Full catalog, not a CLI subset	The Foundry Local CLI is great for quick experiments, but it only surfaces part of what the runtime can run. Flint talks to the official SDK, so you get the broader model catalog (chat, vision, STT, acceleration variants) without writing and maintaining your own wrapper around the service.
Hardware you already paid for	Hardware-aware recommendations and CPU / GPU / NPU variants through the SDK—not guesswork from CLI flags alone.
One local endpoint for many tools	Start a service and point Continue, Cline, OpenAI SDKs, and other clients at http://127.0.0.1:<port>/v1.
More than a single chat tab	Multi-model pool, Model Arena bake-offs, chat + STT, Monitor (resources, access log, audit).
Foundry-native, not a hack	Built on foundry-local-sdk (catalog, download, load, inference, service)—not fragile CLI scraping.
A path to cloud later	Same OpenAI-shaped surface as Azure AI Foundry — local first, cloud profiles planned.

Flint vs Foundry Local CLI

	Foundry Local CLI	Flint
Model surface	Practical subset for common CLI flows	Broader SDK catalog + acceleration-specific variants
Day-to-day UX	Commands, scripts, your own glue	GUI for download/load/pool, chat, audio, arena runs, logs
External tools	You stand up and wire the OpenAI-compatible service yourself	Start service, copy Integrations snippets, manage bind/port
Custom wrapper	Often needed for a full app experience	You don’t—Flint is the maintained control plane on top of the SDK

Use the CLI when you want a terminal-first workflow. Use Flint when you want the full local model surface and a durable UI/endpoint without owning that wrapper.

Built for: developers and power users who want local models and IDE/agent tools on one endpoint; privacy-sensitive or offline-friendly workflows; people evaluating models before committing disk and VRAM.

Not built for: zero-install “ChatGPT clone” installs (see Requirements); fully autonomous agents inside the app (use an agent client against Flint’s endpoint); training or fine-tuning.

---

What’s in the app

• Models — catalog, search/filter, hardware-aware picks, download/load/unload, multi-model pool, update notifications per acceleration track
• Chat — streaming, conversations, personas, system prompts, multi-image vision, host-aware context, optional URL → context
• Audio — mic + file transcription (STT)
• Model Arena — side-by-side prompts, ratings, export
• Monitor — pool, resource gauges, access/audit logs
• Integrations — copy-paste setup for OpenAI-compatible tools
• Diagnostics / Settings — service start/stop, bind/port (Apply & restart), autostart, defaults, shortcuts (?)
• Help — first-run coaching, empty-state guidance, and an About strip

No system Node install is required: release builds bundle their own Node 22 runtime for the Foundry sidecar.

---

Status

0.4.4 — signed installers on the releases page. Pre-1.0: expect breaking changes.

• Windows installers are Authenticode-signed (Azure Trusted Signing); bundles ship the Foundry native cores and a pinned Node 22 runtime, verified in CI.
• In-app updater tracks the latest GitHub release.
• macOS builds are unsigned (no Apple Developer account) — install with the one-liner below, not the DMG, or Gatekeeper will call the app "damaged".

Next: RELEASE_ROADMAP.md · docs/BACKLOG.md · Release notes: CHANGELOG.md

Living plan (docs, help, 0.4, 1.0): docs/PRODUCT_PLAN.md · End-user walkthrough: docs/USER_GUIDE.md

---

Screenshots

Main window

Model selection

Chat

Model Arena

Integrations

Settings

Diagnostics

Audio

Monitor

---

Requirements

End users (installed app)

Requirement	Notes
Windows (primary) or macOS Apple silicon	Intel Mac not supported until Foundry publishes darwin-x64 native cores
Node runtime (sidecar)	Release builds ship a bundled Node 22 binary (Tauri externalBin) for the JS sidecar — PATH Node is not required when packaging is complete. Dev/fallback: Node 22+ on PATH. About shows bundled vs PATH.
Foundry runtime	Bundled — you do not need a separate Foundry CLI for normal use

Developers (building from source)

• Node.js 22+ and npm
• Rust + Cargo (Tauri 2)
• Windows: MSVC + Windows SDK (see docs/DEVELOPMENT.md; build-local.ps1 helps wire cl.exe / SignTool)

---

Quick start

Use a release build

1. Install a build from GitHub Releases when available (or build below).

   • macOS: builds are unsigned, so a browser-downloaded DMG is blocked by Gatekeeper as "damaged". Install with:

     curl -fsSL https://raw.githubusercontent.com/joelst/flint/main/scripts/install-macos.sh | bash

     (Already installed the DMG? xattr -cr /Applications/Flint.app fixes it.)

2. Open Flint (release installers include a bundled Node for the sidecar).
3. Download a small starter model → open Chat.
4. Optional: Diagnostics → Start service, then use Integrations to wire other tools.

If the app cannot start the sidecar, install Node.js 22+ LTS as a fallback or reinstall Flint.

Client URL for tools is always http://127.0.0.1:<port>/v1 (loopback). The bind address in Settings controls what the service listens on and may differ (e.g. 0.0.0.0 for LAN). Use Apply & restart after changing bind/port.

Develop from source

npm install
npm run tauri dev

npm run tauri:build    # runs ensure:foundry + frontend build + package
npm run verify:bundle  # checks Foundry natives in the package
npm run run:built      # launch the release binary without installing

Full scripts, sidecar, and versioning: docs/DEVELOPMENT.md
Signing and release pipeline: docs/RELEASE.md

---

Known limitations

• Node: release builds prefer a bundled Node binary; PATH Node remains a dev/fallback. Long-term 1.0 goal is no end-user Node install at all (bundled or Rust).
• Code signing / updater keys may still be operator-configured for public releases.
• Self-signed installers can trigger SmartScreen / Gatekeeper warnings.
• Audio quality depends on the STT model and runtime.
• Tool calling: models may emit tool_calls; Flint’s chat UI does not execute tools — use an agent client against the local endpoint.
• Unit/contract tests are strong; full UI E2E is still light.

---

Tech stack

• Tauri 2 (Rust + system WebView)
• Svelte 5 + SvelteKit + TypeScript
• foundry-local-sdk via a Node stdio sidecar (sidecar/foundry-sidecar.js)

Architecture principles: FLINT_DESIGN_SPEC.md

---

Documentation

Doc	Audience
docs/PRODUCT_PLAN.md	Next work: docs, help, 0.4, 1.0
docs/README.md	Full doc index
RELEASE_ROADMAP.md	Release scorecards & plans
docs/DEVELOPMENT.md	Build, sidecar, versioning
docs/RELEASE.md	Sign & ship
docs/BACKLOG.md	Deferred follow-ups
CHANGELOG.md	Release notes

Historical plans: docs/archive/.

---

License

MIT

Contributing

See docs/DEVELOPMENT.md. Product direction: docs/PRODUCT_PLAN.md and RELEASE_ROADMAP.md.

---

Local-first by default. Foundry Local underneath. Your hardware, your data, your tools.