Run free OpenCode Zen models, Ollama, LM Studio, OpenRouter and NVIDIA inside Claude Desktop and Codex. Claude keeps its own model names — the model you pick answers, at the effort you choose.
Big Pickle, DeepSeek V4.1 Flash Free, GLM-5.3 Flash Free, Nemotron — they only run inside OpenCode, so Bridge runs them in a private OpenCode agent. No key, no tricks.
Claude Desktop
Claude names. Your models.
Claude still shows Sonnet 5, Opus 5, Haiku 4.5, Fable 5 and Sonnet 4.6. You choose what answers each one — with its own effort, from Low to Max.
Codex
Codex names. Your models.
Codex keeps showing GPT-6-Luna, GPT-5.6-Terra… while the backends you choose answer, each at its own effort. Undo restores your config line-for-line.
How it works
Three steps. Thirty seconds.
01
Pick a model
Every model from OpenCode, Ollama and LM Studio — free, local and agent models clearly badged. New downloads appear on their own.
02
Map & pin effort
Point each Claude model name at any backend and pin Low → Max. Requests snap to levels the model really supports.
03
Launch
Open Claude or Codex on your model. Restore puts everything back exactly — journaled before a single byte changes.
Launchpad
Live library
Pull a model. Watch it arrive.
Bridge fingerprints Ollama, LM Studio and OpenCode every few seconds. ollama pull, an LM Studio download or a newly connected provider slides into the library — no refresh button required.
ollama pull llama4-scout
Models discovered0on a real dev Mac
Free Zen models0via the OpenCode agent
New model detected0after ollama pull
Tests passing0plus live smokes
Recorded session · real output
Codex, running on a free model.
Codex CLI → OpenCode Bridge → OpenRouter’s free Nemotron 3 Super. Codex ran its own shell tools through the Responses ⇄ Messages translation.
codex --profile opencode-bridge exec "Run a shell command to count the lines in app.py…"
Light & dark
Made to live on your desktop.
Glass panels, real provider logos and spring motion — in both themes, synced with macOS. Drag to compare.
Built carefully
Powerful, and politely reversible.
Effort
Low → Max, per model
Pin an effort for each Claude model name or for Codex. Unsupported levels map to the nearest real one — never silently dropped.
Local
Ollama & LM Studio
Native Anthropic endpoints. LM Studio’s server starts and loads your model on launch. Nothing downloads by browsing.
Agent
OpenCode does the work
Agent models run OpenCode’s own tools under your policy — edit & run in your project by default, read-only, or full access.
Security
Loopback only
127.0.0.1, Host checks, a random 256-bit token, and a password-protected OpenCode server. Keys stay in OpenCode.
Recovery
Journaled, exact undo
Claude and Codex edits are journaled before any write. Restore detects outside edits and refuses to clobber them.
CLI
Headless when you want
init, launch, status, restore — everything the app does, scriptable.
From source
Prefer the terminal?
Three commands. Node 22+ and the OpenCode CLI are all you need.
zsh
$ git clone https://github.com/satiricalguru/Opencode-bridge.git
$ cd Opencode-bridge && npm install
$ npm run app
v0.4.2 · Apple Silicon
Download OpenCode Bridge
For macOS on Apple Silicon. Free and open source under MIT.
