Mr Rawbot
GitHub

Design the work. Delegate the build. Stay in control.

Mr Rawbot is a local command center for AI-assisted software work. Shape product ideas into specs and prototypes, hand implementation to Claude Code, Codex, or Ollama Cloud, and review every run without scattering context across chat windows, terminals, and browser tabs.

Download for macOS View source on GitHub
PlanClaude Code BuildOllama Cloud ReviewCodex Review loops back until the work is approved.
Mr Rawbot running a Plan → Execute → Review workflow with per-step status, tokens, cost, and duration
Overview

One loop from product direction to finished work

Mr Rawbot gives the messy parts of AI-assisted development a single home: product direction, generated prototypes, implementation runs, review evidence, usage, and handoff context. It is where you decide what should be built and keep track of what the agents did.

shape · delegate · review
Abstract routing visual showing product design and agent paths converging into one local workspace
Local workspaceProjects, conversations, artifacts, runs, and provider settings live on your machine.
Accountable runsEach agent handoff carries progress, logs, usage, cost, and final changes.
Design-to-buildShape the product first, preview the prototype, then attach it to implementation.
Orchestration

Give the right agent the right job

Some tasks need one strong model. Others need a planner, a builder, and a reviewer. Saved workflows let each step choose its provider, model, role, instructions, effort, and loop rule — so Claude, Ollama, and Codex can work as specialists instead of interchangeable chat tabs.

  • Start with a single-agent run when the task is straightforward
  • Use plan–build–review workflows when the work needs checks and handoffs
  • Create custom workflows with provider-specific roles and completion rules
The workflow editor showing a Plan → Execute → Review pipeline with per-step provider and model choices
Review

See the result before you accept it

Agent output is only useful if you can inspect it quickly. Mr Rawbot keeps the generated changes, affected files, branch state, and pull request context next to the conversation, so review stays part of the workflow instead of a separate cleanup step.

  • Browse the selected project and inspect generated edits in context
  • Compare before-and-after changes inline when you need detail
  • Track branch state, pull requests, commits, pushes, and merge readiness
Review panel showing changed files and an inline before-and-after comparison
Runs

Know what happened, what it cost, and why

Every agent run leaves an audit trail: status, logs, duration, model, reasoning effort, token usage, and estimated cost. When a task goes sideways, the evidence stays attached to the run instead of disappearing into temporary console output.

  • Live step timeline while the agent works
  • Input, cached, output, and reasoning-token accounting where providers expose it
  • Run history saved with the conversation that produced it
A single-agent Claude run with its log timeline, token counts, and cost
Product design

Shape the product before the code changes

Product Design is its own thread mode, not just another workflow. The Product Specialist turns goals into specs and build prompts; the Product/UI Designer creates or revises high-fidelity HTML/CSS prototypes in an app-managed workspace. Use it to make the product decision first, then attach the artifact to implementation.

  • Specs with user stories, acceptance criteria, and build-ready scope
  • Multi-page HTML/CSS prototypes with embedded preview
  • Build prompts that carry the visual and product decisions into implementation
Artifacts gallery with a rendered email prototype, a spec, and a build prompt
Providers

Bring the agents you already trust

Claude Code

Use your local claude login for planning, implementation, tests, and broad codebase work through the Anthropic Agent SDK.

claude-opus-4-8claude-fable-5claude-sonnet-5

Codex

Use your existing codex login for focused builds, review passes, and app-managed Codex runs.

gpt-5.5gpt-5.4gpt-5.4-mini

Ollama Cloud

Run an Ollama Cloud coding agent with project tools and local shell access, configured from Settings.

qwen3-coder:480b-clouddeepseek-v3.1
Built-in flows

Choose an operating mode

Claude Code
One Claude agent owns the task from request to result.
Codex
One Codex run handles implementation, debugging, or review.
Ollama Cloud
One Ollama Cloud agent works with project tools and shell access.
Plan → Build
Claude defines the work; Ollama carries it through with completion checks.
Heavy Plan → Build
Claude spends more time on architecture, risks, and edge cases before the build step.
Plan → Execute → Review
Claude frames the work; Ollama builds; Codex reviews and loops until APPROVE.
Ollama Plan → Codex Build
Ollama drafts the approach; Codex performs the build.
Local-first

Your machine is the control plane

Mr Rawbot runs locally with no hosted backend, no product account, no telemetry, and no .env file. Provider calls still go through the Claude, Codex, and Ollama accounts you configure; Mr Rawbot keeps the orchestration state, artifacts, runs, and settings on your machine.

  • Threads, runs, artifacts, and settings are persisted locally in SQLite
  • CLI paths are auto-detected; provider secrets are encrypted at rest
  • Project discovery skips VM mounts and container filesystems, and manual folder paths are supported
Settings dialog showing provider configuration with auto-detected CLI paths and available models
Install

Install the desktop app or run it locally

Download the app

macOS builds ship as .dmg artifacts for Apple Silicon and Intel. The packaged app uses your system Node and local provider logins; because it is unsigned, clear quarantine once after installing:

# after dragging it to /Applications
xattr -dr com.apple.quarantine "/Applications/Mr Rawbot.app"

Get the latest release →

Run from source

Use the source build on any platform with Node.js 24 available. First-run setup happens inside the app under Settings → Providers.

git clone https://github.com/marcospcury/mrrawbot
cd mrrawbot && npm install
npm run dev    # web app at localhost:5173
npm run app    # native Electron window