About Services Tools Blog Contact us

Local Coding Agent

A coding assistant that reads and writes a private codebase, runs its checks, and works with a model on the same Mac so the code stays on the machine.

  • Agentic Workflows
  • Professional services
  • Agent skill
  • Product
  • Local/private
  • 2026

The problem

Small code changes still need a person to open the project, run the checks and report back. An assistant that can only call an online service cannot be used on a private codebase.

What it does

The assistant lists, searches, reads and writes files, runs the project's tests and style checks, and can run a command with a time limit. The model runs on the same machine through Ollama, with a single-prompt mode for scripts and a repeat mode for unattended passes. It should only run on a trusted machine.

Outcomes

  • A small change comes back as a diff and a test result without a person driving the editor.
  • Private code stays with the model on the local machine.
  • A script can run a single prompt.
  • A transcript records what the agent read and changed.

Simplified mockup

Interactive preview with sample data. Nothing is sent to a server.

Dispatch labels loop · every 30 minutes

Repository private · this machine
dispatch/folder
labels.phpto edit
labels_test.pytest
readme.mdleave
Task

Rename the dispatch label from “Ship now” to “Send today”. Keep the test names. Do not touch the readme.

Compact 4B. Short context, enough for one label.

Balanced 9B. Reads the label and the test before writing.

Run no shell is started here
Idle. The loop chip is a schedule, not a live timer.

labels.php

No diff yet. Run the rename from the session tab.

Session transcript

Yesterday · collection wording

Headless prompt. Two files. Local model. Timed out once, then passed.

passed
Monday · lint only

No writes. Stopped at the timeout before a second lint.

timeout
Unattended loop every 30 minutes

The next pass repeats this prompt only when the working tree is clean. A failing test or a timeout stops the loop. Nothing here starts a real timer.

How it works

  1. Point the session at a codebase and a local model.
  2. Read the task and search the code before editing.
  3. Write files and show the changes.
  4. Run tests and style checks, and stop if they take too long.
  5. Keep a record of the session, and optionally repeat the prompt on a timer.
  6. Add a reusable tool only when the session is allowed to.

Pipeline

flowchart LR
  A[Task] --> B[Code search]
  B --> C[File edit]
  C --> D[Diff]
  D --> E[Tests]
  E --> F[Transcript]