For Claude Code · Claude Desktop · Codex CLI · OpenCode

Bridge1

Runs your coding agent on your own local models without losing the reliability you'd get from a frontier API — real context handling, models kept warm between turns, one setup for every client.

curl -fsSL https://bridge1.ai/install.sh | sh

macOS · installs to Applications · read the script · or download the .dmg

See how it works

How it works

Three steps, all on your own machine.

1

Install Bridge1

One local gateway process runs on your machine — nothing to host, nothing sent to a third-party server just to route a request.

2

Add your models

Pick from the catalog below in one dashboard — llama.cpp or MLX, sized and verified for what your machine can actually run.

3

Point your client at it

Copy the setup snippet for Claude Code, Claude Desktop, Codex CLI, or OpenCode from the Gateway panel. One paste, done.

Local model catalog

A curated set of models Bridge1 has actually tuned and verified per entry — real context length, real RAM/VRAM needs, not a spec-sheet guess.

llama.cpp

Qwen3.6-27BQwen3.6-35B-A3BQwen3.8-27BMuse Glimmer 30BOrnith-1.5-9BOrnith-1.5-35B-A3B

MLX (Apple Silicon)

Gemma 4 26B-A4BQwen3.6-27BQwen3.6-35B-A3BQwen3.8-27BQwen3-VL-2B-InstructTernary-Bonsai-27B

The catalog grows over time — each new entry gets the same real per-model verification before it ships, never added sight-unseen.

Supported clients

The four marked "Ready" are confirmed working end-to-end. Hermes Agent and Pi.dev are real, wired integrations with a bit more setup — not yet as polished.

Claude CodeReady
Claude DesktopReady
Codex CLIReady
OpenCodeReady
Hermes AgentEarly support
Pi.devEarly support

Any other OpenAI-compatible client can also point at Bridge1 directly.

Download

Bridge1 asks for an invite code the first time you run it (each code activates one device) — a friend who already has Bridge1 can share one with you from inside the app.

curl -fsSL https://bridge1.ai/install.sh | sh

Recommended — no macOS security warning. Read the script first.

Download the .dmg instead

macOS today — Windows and Linux are architected in but not yet built or tested.

Memory: 24 GB minimum on Apple Silicon, 32 GB or more recommended. Bridge1 runs models on your own machine, so memory is the real limit — an 8 GB or 16 GB Mac cannot run any chat model in the catalog. Bridge1 shows each model's measured RAM requirement inside the app before you download it. Intel Macs are untested.

Recommended: install from the terminal — no macOS warning at all.
curl -fsSL https://bridge1.ai/install.sh | sh
macOS only flags apps downloaded by a browser. Installing this way skips that entirely, and the script is plain text you can read first at bridge1.ai/install.sh.

If you download the .dmg instead, macOS will warn you the first time you open Bridge1. It says Apple "could not verify" Bridge1 — that is because this early-access build is not yet notarized by Apple, not because anything is wrong with it. Notarization is in progress.

The dialog's blue button is "Move to Trash". Don't click it. Click Done, then follow the three steps below — it takes about twenty seconds, once.
The 20-second fix
  1. Double-click Bridge1 once. It will be blocked — that is expected. Click Done, never Move to Trash.
  2. Open System Settings → Privacy & Security, scroll to "Bridge1" was blocked..., click Open Anyway, then confirm.
  3. On macOS versions before Sequoia, right-click Bridge1.app → Open → Open also works.
  4. Terminal fallback: xattr -d com.apple.quarantine /Applications/Bridge1.app

Don't have an invite code? Ask someone who already runs Bridge1 — every activated install can share a few codes each week.
Terms·Privacy