Existing behavior stays
No changes to UI, keybindings, agents, prompts, tools, MCP, permissions, sessions, or retry.
OpenCode v1.18.30 × MLXBar
Local Mac AI, inside OpenCode.
OCode mlx-bar is a minimal fork that leaves OpenCode's UI, agents, tools, MCP, and sessions intact, adding only MLXBar as a first-class provider.
Free · MIT License · ad-hoc signed · not notarized by Apple
OpenCodeUI · Agent · Tools
↓
provider / mlxbarminimal adapter
↓
MLXBar :11435OpenAI compatible
↓
Apple Siliconlocal inference
● connected · models discovered · stream ready
Design contract
One provider. Nothing else rewritten.
A small delta is the quality bar. When MLXBar is not selected, observable behavior is intended to match OpenCode v1.18.30.
Native Auto Compaction
MLXBar's effective context and safe-input limits feed the existing OpenCode model metadata and compaction policy.
Starts without MLXBar
Discovery is brief and bounded. An offline server cannot take down OCode or another provider.
Provider
Right where OpenCode already lists models.
There is no custom wizard or separate screen. Start MLXBar, then use the provider and model selectors you already know.
Dynamic discovery
Loads the current model inventory from GET /v1/models during startup.
Streaming chat
Uses the existing OpenAI-compatible SDK path for streaming responses and tool calling.
Local or LAN
Use the default localhost endpoint or supply a LAN base URL and token through normal provider settings.
No capability guessing
Vision, reasoning, and token limits are enabled only from explicit metadata—not model names.
Get started
Start MLXBar. Pick a model.
The default local setup needs no extra configuration. Use standard OpenCode provider settings only for token authentication or LAN access.
- Start MLXBarLoad a model and enable its OpenAI-compatible API.
- Open OCode mlx-barIt briefly probes
/v1/modelsduring startup. - Select an MLXBar modelChoose
mlxbar/<model>in the existing model selector.
{
"provider": {
"mlxbar": {
"options": {
"baseURL": "http://127.0.0.1:11435/v1",
"apiKey": "YOUR_MLXBAR_TOKEN"
}
}
}
}Safety by scope
Bounded, isolated, and explicit.
Host memory pressure and context overflow are different conditions, even for a local provider.
Ambiguous memory pressure stays intact
It maps to context overflow only when MLXBar explicitly marks it context-recoverable.
No invented limits
When a safe-input limit is absent, the adapter never substitutes a percentage heuristic.
Memory boundaries
Discovery is capped at 4 MiB and 10,000 models.
No secret logging
The adapter does not log API keys, authorization headers, prompts, or tool results.
Release facts
First release, clearly defined.
OCode mlx-bar is free and open source. OpenCode's original MIT license and copyright notice are preserved.
The macOS build is ad-hoc signed and not notarized by Apple. First launch may require manual approval in System Settings › Privacy & Security.