Earth Link Network Co., Ltd.

local-commander

A development orchestrator that puts a local LLM in the decision-maker's seat. It sorts tasks by difficulty, inserts human approval only where needed, and automates everything from implementation and review to merge.

local-commander is an AI development orchestrator that puts a local LLM in charge at runtime. It has been validated on a MacBook Pro M3 with LM Studio, and is built as a prototype app intended to move to always-on operation on an NVIDIA DGX Spark. It classifies development tasks and records the Implementation Brief, the prompt for Codex, the evaluation of Codex's results, and token usage (real API usage plus estimates).

Codex CLI and Claude Code are never put in charge. Codex CLI is an external agent that is only called when the local LLM judges a task `cloud-needed`, the user explicitly passes `--run-codex`, and the human approval gate has been passed.

Key Features

Task classification and Implementation Brief generation

The local LLM classifies tasks as local-ok / cloud-needed / spec-missing and generates the corresponding Implementation Brief.

External agent handoff with a human approval gate

Codex CLI (or Claude Code, when --agent claude is specified) is only called when a task is cloud-needed, the user explicitly passes --run-codex, and human approval (--human-approved) has been given.

Token usage logging and cost aggregation

Records character counts and estimated token counts (chars ÷ 3.5) for each step, and also saves actual token counts when an OpenAI-compatible API returns usage. npm run cost aggregates by model and by runId.

Live control panel (Web UI)

The persistent Web UI started with npm run server (default http://localhost:4317) gives a single view of the job list, the approval inbox, and cost.

Claude Code MCP integration (eln-judge)

Offloads only the local LLM's judgment as a stdio MCP tool. The shim reads file contents locally, and the tool's return value contains only the judgment and a summary (v1 is read-only).

Asana / GitHub integration and classification benchmarking

Supports TaskSources that pull tasks from the Asana API and the GitHub API, and npm run benchmark measures classification accuracy, safety, and speed across multiple models (M1-M6).

Internal use

Back to the product list

For inquiries about local-commander, please use "Contact" in the header.

local-commander