← back

the conductor

Maestro

A tiny, local-first AI orchestrator and agent harness for Apple Silicon. It chains local models into multi-step pipelines with a web UI, a drag-and-drop visual canvas, and a ReAct subagent loop — and the core runs on the Python standard library alone.

Python stdlib Apple Silicon oMLX visual canvas

Maestro conducts local models. An orchestrator is a DAG of steps; each step runs an agent — a harness plus a model — that can reason, call tools, and hand its output downstream. One window, two views: Build is a hand-rolled SVG canvas with no UI library and no build step, and Run view is a monitor — a progress bar over a resizable Log · Thinking dock where the model's reasoning streams beside the log while nodes light up.

It handles the shapes real workflows actually have: bounded loops that stay out of the DAG, evaluator steps that route on a model's verdict, pure-code branches, and human approval gates that pause a run until you click Approve — right on the canvas node. The shipped Nerd orchestrator builds whole apps this way: an architect files a plan into an append-only to-do, workers drain it one file at a time, and a reviewer re-queues fixes until the build passes.

The core has no pip dependencies at all. Inference is served by oMLX, the native MLX server, over HTTP — so models load once, in one place — and every Python skill runs as an isolated child process that can't take the server down.

highlights

canvas

visual pipeline editor

Drag nodes for agents, tools, routing, evaluators, gates and transforms — or another whole orchestrator as a single step; drag a dot to wire a dependency or a loop link.

agents

Agent = harness + model

System prompt, memories, skills, and bounded tool loops — with context and observation caps so local prefill cost stays predictable.

flow

loops without cycles

The graph stays acyclic; bounded loop links and evaluator verdict routing handle retries, chapter loops, and build-review cycles.

runs

walk away mid-run

Runs are detached and concurrent: close the tab, come back from another browser, follow any live one — thinking pane included. Any recorded run resumes from any step.

loop

the Nerd build loop

Architect plans an append-only to-do, workers implement one task per turn, a reviewer re-queues fixes until the build passes.

reliability

fails loud, not silent

Every run is validated before any model call, an unreachable backend raises instead of forging a final answer, and Stop cancels the model call itself — not just the loop around it.

under the hood