A desktop app that runs a local LLM gateway on your Mac/Windows/Linux box and exposes it as an OpenAI-compatible endpoint, so Cursor, Cline, Cherry Studio, or anything else that talks /v1/chat/completions can hit one URL and reach 37 providers behind it.
Read this in 简体中文.
Prism wraps the one-hub gateway as a Wails (Go + React) desktop app. The gateway runs in-process — no Docker, no separate docker-compose.yml, no port hunting. The app also bundles a cloudflared binary, so one click gives you a public *.trycloudflare.com URL when you want to share the endpoint with a teammate or a remote agent.
What you get:
- One OpenAI-compatible HTTP endpoint, local or temporarily public.
- 37 providers behind it: OpenAI, Anthropic, Gemini, Bedrock, Vertex AI, DeepSeek, Groq, Zhipu, Moonshot, …
- A real desktop UI for managing channels, tokens, request logs, and full request/response traces.
- A copy-paste snippet generator that produces ready-to-use Cursor / Cline / Cherry Studio configs.
It is one binary. No external services required.
I wanted to use Cursor and Cline against several providers (some only reachable through a proxy, some only available regionally) without juggling N config files, and without running another always-on Docker container on my laptop. one-hub solves the routing problem; Prism is the part that makes one-hub usable as a daily-driver desktop tool: configuration UI that doesn't fight you, a public URL on demand, and full traffic capture so when a provider returns a weird 400 you can see exactly what went over the wire.
Grab a build from Releases:
| Platform | File |
|---|---|
| macOS (Intel + Apple Silicon) | Prism-vX.Y.Z-darwin-universal.dmg |
| Windows amd64 | Prism-vX.Y.Z-windows-amd64-installer.exe (NSIS) or .zip (portable) |
| Linux amd64 | Prism-vX.Y.Z-linux-amd64.tar.gz |
Releases are unsigned. First-run notes:
# macOS Gatekeeper
xattr -dr com.apple.quarantine /Applications/Prism.appOn Windows, SmartScreen will warn — "More info" → "Run anyway". On Linux, you need GTK 3 and libwebkit2gtk-4.1-0 (default on Ubuntu 22.04+, Fedora 39+, Debian 12+).
- Launch Prism. The onboarding card walks you through it.
- Channels → New — pick a provider, paste an API key.
- Bedrock key format:
region|AccessKeyID|SecretAccessKey. - For providers blocked from your network (e.g. Anthropic / Gemini direct from China), put your local proxy in Outbound proxy, e.g.
http://127.0.0.1:7890.
- Bedrock key format:
- Tokens → New — give it a name. "Unlimited quota" is fine for single-user setups.
- Dashboard → Start tunnel — Prism launches
cloudflaredand shows you a public URL within a few seconds. (Or skip the tunnel and usehttp://127.0.0.1:<port>from the same machine.) - On the Tokens page, click the snippet icon, pick your client, and paste the resulting JSON into Cursor / Cline / Cherry Studio.
- Embedded one-hub. No separate process. The Go binary imports the gateway as a library, registers its routes on a
gin.Engine, and listens on a chosen local port. SQLite by default; switching to MySQL / Postgres still works (SQL_DSNenv var). - Tunnel, with no Cloudflare account.
cloudflaredis bundled per-platform (go:embed) and run as a child process; the app parses its log to surface the public URL in the UI. No tunnel-config file, no zero-trust setup needed. - Traces. A small Gin middleware tees the request body and the response body (including SSE streams) into a SQLite table. When a provider returns
tools.0.custom.name: Field requiredand the wrapper says "Provider API error", the actual upstream payload is one click away in the Logs → Traces view. - macOS menu bar. A native
NSStatusItem(CGO + Objective-C, not a third-party systray library — those collide with Wails' ownAppDelegate). Closing the window keeps the process running, drops the Dock icon, and leaves the menu bar item with Start/Stop tunnel and Copy URL actions. - Theming and i18n. Light/dark themes, English/中文.
Vendored at backend/core/. Notable diffs from upstream (MartialBE/one-api):
core/providers/claude/chat.go— when a client like Cursor sends Anthropic-shaped tools to the OpenAI-compatible endpoint, the upstream rejected them withtools.0.custom.name: Field required. Patched to settype: "custom"and to threadinput_schemathrough correctly.core/providers/gemini/chat.go— top-levelsystemprompts and array-style system messages now both aggregate into Gemini's expectedsystemInstructionfield.core/common/gin.go—ErrorWrapperno longer swallows the underlying network error behind the generic "请求上游地址失败". The originaldial tcp …/i/o timeoutis appended to the response body so it shows up in Traces.core/providers/bedrock/*— Anthropic-on-Bedrock tool schema fixed to match the patches above.backend/prism/capture/middleware.go— new. Non-destructive body tee with agin.ResponseWriterproxy that handles streaming and stores everything toprism_traces.
If a one-hub upstream PR lands that obsoletes any of these, the diff goes away.
# prerequisites: Go 1.24+, Node 20+, pnpm 9+
go install github.com/wailsapp/wails/v2/cmd/wails@latest
# fetch the cloudflared binaries (not in git — too big, fetched at build time)
./scripts/fetch_cloudflared.sh
wails dev # hot reload
wails build -platform darwin/universal # production build
wails build -platform windows/amd64 -nsis
wails build -platform linux/amd64 -tags webkit2_41The artefact lands in build/bin/.
User config lives at:
- macOS:
~/Library/Application Support/Prism/ - Windows:
%APPDATA%/Prism/ - Linux:
~/.config/Prism/
…and contains prism.yaml (runtime config; secrets auto-generated on first launch), prism.db (SQLite), and logs/.
Pushing a v*.*.* tag triggers .github/workflows/release.yml, which builds the matrix on macOS / Windows / Linux runners and uploads the artefacts to the corresponding GitHub Release. To produce just one platform on demand:
gh workflow run release.yml --ref main \
-f tag=v0.1.1 \
-f platforms=linux-amd64Patches are welcome. Conventions in CONTRIBUTING.md. Please don't paste real API keys into bug reports — Prism's Traces view is private to your machine, so screenshots from there are usually safer than copy-pasted JSON.
Security issues: see SECURITY.md.
Why AGPL: Prism is a network service — it speaks an OpenAI-compatible HTTP API to remote clients. AGPL covers exactly that case: anyone running a modified Prism over a network has to make their source available to its users. It also matches the copyleft expectations of the embedded one-hub (GPL-3.0), which permits redistribution under stricter copyleft.
- one-hub — the gateway core. Prism is roughly "one-hub minus the docker-compose, plus a desktop UI".
- Wails — the Go-React desktop framework that makes embedding Go gateway code into a native window not painful.
- cloudflared — the public-URL story.