BRIDGE1 Full Context
Full Context means the largest window your Mac’s memory can hold, up to each model’s maximum; client limits still apply.
Serving path →Bridge1 is a local inference engine for vibe coding, built on MLX and llama.cpp.
Local LLM for Vibe Coding. Nothing to Tune.
Local models make you the engineer: engine flags, memory budgets, context sizes, per-tool config. Bridge1 handles all of it, so you pick a model and vibe code.
Works with OpenCode, Codex, Claude Code and Claude Desktop on Mac.
24 GB RAM minimum, 32 GB recommended · Check your fit
OUR MISSION
Local AI should feel like the cloud.
Full context, fast response, nothing to tune, and your code never leaves your computer.
We're building Bridge1 so running AI on your own machine feels as effortless as the cloud: full context, managed memory, nothing to tune.
With a local model on Bridge1, your prompts and outputs stay on your computer, and Dojoit runs no server that receives them.
How Bridge1 handles data →01 / SPEED YOU FEEL
Long prompts cost prefill. Reuse the matching prefix.
A 21.7 GB model loaded in 8–9 s in one past test. A long prompt still has to be read once, and a reasoning model may think before its first visible answer. On MLX, Bridge1 can reuse an exact prompt prefix from a cache on disk, even after the app restarts. We have verified that on two models, not yet on the default Ornith-35B.
01 · LOAD8–9 sto load a 21.7 GB model · Ornith-35B Q4_K_M · llama.cpp · load count not recorded 02 · WARM TTFT~50 msto the first token on a same-chat follow-up · 49–51 ms over three requests after a 1.5 s first request · Ornith-35B Q4_K_M · llama.cpp · prompt length not recorded 03 · AFTER A RESTART99.7%of prompt tokens reused on an identical resend after quitting and reopening the app · Qwen3.8 27B and Qwen3-VL-2B · MLX · one run each 04 · SECOND TURN22.8xrecorded second-turn ratio, 359.7 s to 15.8 s, in one 28.7K-token session · Qwen3.8 27B 4-bit · MLX
The Bridge1 figures on this strip come from one Apple M1 Max with 64 GB, on builds before the 0.7.3 release, and most are a single run. Other Macs are not measured. A cold time to first token on the released build is not yet measured.
02 / SEE IT WORK
From download to first reply, on your own Mac.
A scripted walkthrough of the flow: get a model, route it, connect OpenCode, start coding.
Illustrative walkthrough. Scripted, not a live run.
The scripted reply appears beside an illustrative Bridge1 Observability view: an active session, Live Log entries and moving throughput and context indicators.
- Pick a model
Download a model in Local Models. The scripted download ends with the model ready to use.
- Serve it as Route1
Open Manage on the model and turn on Route1. Clients point at Route1 once; the model behind it can change without editing client settings.
- Open the Gateway
Open Gateway and select OpenCode. Copy the provider ID, display name, base URL and masked API key into the custom provider form.
- Add Bridge1 to OpenCode
In OpenCode, open Settings, Providers and Custom provider. Add Bridge1 using the gateway base URL, API key and route1 model ID.
- Start a session
Select route1 in an OpenCode session and send a coding prompt.
- Watch it work
The scripted reply appears beside an illustrative Bridge1 Observability view: an active session, Live Log entries and moving throughput and context indicators.
03 / WHAT BRIDGE1 DOES
Local inference for coding, with the engine and memory managed.
Running a local model for OpenCode, Pi.dev or Codex should not mean configuring model servers, runtime flags and a different endpoint for each client.
Bridge1 runs the model engines, checks memory before each load, and keeps one local address for your clients.
Explore Bridge1 features for local coding.
Low memory
Low memory starts with a fit check: memory is checked before every model load. The catalog separates measured memory rows from estimates.
Memory rules →Measured speed
Fast in one historical chat: Qwen3.8-27B 4-bit MLX on M1 Max 64 GB took 15.8 s on turn 2 versus 359.7 s cold on turn 1, reusing 28,722 tokens (one run).
Measured comparisons →Drafter Fusion
Drafter Fusion uses DFlash or DFlash2 checkpoints, MTP heads, n-gram lookup or no drafter, depending on the catalog entry. Read the per-model results and limits.
Explore Drafter Fusion →Method credits →04 / VERIFIED AND VISIBLE
Every model file is checked. Every path is reported.
Avoid mismatched model files and engine builds: each download is pinned to a source revision, a size and a SHA-256 hash. The Live Log and dashboard show which model, engine, context and speed path actually ran.
Downloads you can check
Model files and engine binaries are fixed to a revision and hash. A mismatch stops the load.
Show what ran
The Live Log and dashboard show the model, the engine, the context window and each speed path as engaged, declined or unavailable.
Speed depends on the model
The dashboard identifies the active prediction path for each model. Read the story →
05 / HOW IT WORKS
Install. Add a model. Connect your agent.
- 1Install and activate
Install the macOS app. Enter an invite code if your early-access activation asks for one.
- 2Add a model
Choose a model from the catalog. Bridge1 downloads the pinned files and prepares its engine.
- 3Copy your client setup
The Gateway panel generates the setup for your client, so its API format and model name do not become another config to maintain by hand. On macOS, Claude Code, Codex, OpenCode and Pi can also be started with bridge1 launch.
- 4Switch models when you need to
Keep the client pointed at Route1. Optionally measure a model on your Mac first with Calibrate.
06 / EVIDENCE
Measured on one Mac. Shown with its conditions.
Every Bridge1 runtime figure here was measured on an Apple M1 Max with 64 GB. Runtime measurements and external model-quality scores are kept separate.
TWO-TURN SESSION · ONE RUN22.8x
Turn 2 took 15.8 s after turn 1 took 359.7 s, with the whole 28,722-token prefix reused.
- Turn 1
- 359.7 s · no cache
- Turn 2
- 15.8 s · 28,722 tokens reused
- Model
- Qwen3.8 27B, 4-bit MLX
- Hardware
- Apple M1 Max · 64 GB
One run. A single session is not a promise for other models, prompts or machines.
40 SMALL CODING TASKS · CLAUDE CODE40 of 40
Ornith 1.5 35B-A3B passed every task in our tests. The reference model, Qwen3.8 27B, passed 39 of 40.
- Qwen3.6 35B-A3B
- 38 of 40
- Qwen3.5 4B (10 tasks)
- 9 of 10
- Tasks
- Small repositories, hidden unit tests
- Limit
- Claude Code only · not a public benchmark
Full data and limits are on the Evidence page.
07 / CURATED MODELS
Local coding models, with quantization and decode evidence.
Bridge1 recommends Ornith 1.5 35B-A3B for agentic coding and lists Editor's Choice models for each engine. Every profile shows its size on disk, memory and measured speed.
Ornith 1.5 35B-A3B
OptiQ 4-bit · 81.2 tok/s decode (conditions not recorded) · 79.0 SWE-bench Verified (base weights)
Qwen3.6 35B-A3B
mxfp4 · 94.8–98.6 tok/s short decode · 47.1 tok/s at depth
Ornith 1.5 9B
Q4_K_M · 42.0–44.6 tok/s decode (conditions not recorded) · 70.6 SWE-bench Verified (base weights)
Speeds are measured on an Apple M1 Max, 64 GB, as catalog figures from 2026-10-09. Other Macs are not measured yet.
08 / CLIENTS
Works with OpenCode, Codex, Claude Code and Claude Desktop
Early support: Pi.dev and Hermes Agent.
OpenCode, Codex, Claude Code and Claude Desktop are tested end to end on Bridge1.
OPENCODE · READYPI.DEV · EARLY SUPPORTCODEX · READYHERMES AGENT · EARLY SUPPORTCLAUDE CODE · READYCLAUDE DESKTOP · READY
Setup and known limits for each client → 09 / AVAILABILITY
Check your Mac before you install.
Bridge1 runs on macOS today. Apple Silicon is recommended. Each model lists its own memory and disk needs, and Bridge1 checks memory again before every load.
- PLATFORMmacOS early access · Apple Silicon recommended
- MEMORY24 GB minimum · 32 GB or more recommended
- WINDOWS AND LINUXNot available
LOCAL AGENTIC CODING
Get early access.
Compare local inference designs →
Bridge1 questions
Do I need to configure anything to run a local LLM for coding?
No engine tuning is required for a catalog model. Bridge1 manages MLX and llama.cpp, checks memory before each load and sizes context to fit. Choose a model, then copy the setup for your client instead of maintaining engine flags and per-tool model settings.
See the inference engineWhat is Bridge1 for local vibe coding?
Bridge1 is a local inference engine for vibe coding, built on MLX and llama.cpp. It runs a Large Language Model (LLM) on your Mac and manages engine settings, memory fit and context, so a coding session starts with a model choice rather than runtime tuning.
How does a coding request reach my local model?
Your client sends a request to one local gateway. Bridge1 adapts its API format and tool calls, dispatches to the chosen model on MLX or llama.cpp, then streams the reply back. A memory plan admits the model before loading, avoiding a guess about whether weights and context fit.
Which coding clients can use Bridge1 local inference?
OpenCode, Codex, Claude Code and Claude Desktop are tested end to end; Pi.dev and Hermes Agent have early support. One local address and generated per-client setup reduce separate configuration work. API compatibility alone does not remove client watchdog, tool-call or compaction limits.
Check client-specific limitsDoes local LLM inference send my prompts to Dojoit?
Local model prompts and outputs stay on your computer. Bridge1 still contacts services for activation, catalog updates and update checks, plus waitlist or feedback you choose to send. Running inference locally removes the model round trip; it does not make every app feature offline.
What memory does my Mac need for local coding with Bridge1?
The macOS offer needs at least 24 GB of memory; 32 GB or more is recommended. Apple Silicon is recommended and Intel is untested. Model weights are only part of the memory bill: context and engine overhead also count, so Bridge1 checks fit again before loading.
Check unified memory requirementsHow do I get invite-only Bridge1 early access?
Request an invite on the Download page, or redeem a code you already have. The macOS app prepares the chosen model and engine, so your next local coding session can use generated client setup. Early access has no published acceptance deadline or guaranteed place.
Get macOS early accessIs Bridge1 free? Is it open source?
The Bridge1 local gateway is currently free of charge, and no account or payment is required to use it. Bridge1 is not open source today; we may open-source it in the future, but we have made no commitment. See the Terms for details.
How is Bridge1 different from Ollama or LM Studio?
Ollama and LM Studio are good tools for broad model choice. Bridge1 is a local inference engine for vibe coding, built on MLX and llama.cpp. It manages memory-fit checks, context, Drafter Fusion, cache reuse and client handling. Its curated catalog simplifies setup, but you cannot load arbitrary models.
Bridge1 vs Ollama vs LM Studio