ashlr-workbench
Local agent workbench bundling OpenHands, Goose, Aider, and ashlrcode against one local LLM, with ashlr-plugin MCP servers pre-wired.
ashlr-workbench
Your local coding-agent HQ — four agents, one local LLM, ten MCP servers, zero cloud dependencies.
What this is
A curated, fully-local toolbox that wires four open-source coding agents
(OpenHands, Goose, Aider, ashlrcode) to a single LM Studio model
(Qwen3-Coder-30B) and the same set of ten ashlr-plugin MCP servers. The
workbench supplies one CLI (aw), one healthcheck, and one update path so the
whole stack feels cohesive instead of four loose tools you have to babysit. No
data leaves the machine unless you opt into a cloud fallback.
Quick start
# 1. Clone the supporting plugin (provides the 10 MCP servers)
git clone https://github.com/ashlrai/ashlr-plugin ~/Desktop/ashlr-plugin
cd ~/Desktop/ashlr-plugin && bun install
# 2. Clone the workbench
git clone <this-repo-url> ~/Desktop/ashlr-workbench
cd ~/Desktop/ashlr-workbench
cp .env.example .env # then edit
# 3. Start LM Studio, load qwen/qwen3-coder-30b, click "Start Server"
# 4. Verify
./bin/aw doctor
# 5. Use
./bin/aw start aider # interactive session in cwd
./bin/aw start openhands # autonomous Docker-based agent
To put aw on your PATH (review first):
ln -sf ~/Desktop/ashlr-workbench/bin/aw /usr/local/bin/aw
The agent lineup
| Agent | Form | Best for | Cost |
|---|---|---|---|
| OpenHands | Docker daemon + GUI | Autonomous multi-step PRs, browser-driven tasks | Heavy RAM |
| Goose | Native Rust CLI | Fast tool-using sessions, smart-approve loop | Light |
| Aider | Python CLI | Surgical file-by-file refactors with explicit diffs | Light |
| ashlrcode | npm/Bun CLI (Mason's) | Personal day-to-day work, hooks-aware, cloud fallback | Light |
Pick the one that matches the shape of the work — see "Usage examples" below.
Architecture
┌─────────────────────────────────────────────┐
│ ashlr-workbench │
│ │
│ bin/aw → scripts/{start,health,update} │
└─────────────────────────────────────────────┘
│
┌───────────────────┬───────┴────────┬──────────────────┐
▼ ▼ ▼ ▼
┌──────────┐ ┌────────┐ ┌────────┐ ┌──────────┐
│ OpenHands│ │ Goose │ │ Aider │ │ashlrcode │
│ (docker) │ │ (rust) │ │(python)│ │ (bun) │
└────┬─────┘ └───┬────┘ └───┬────┘ └────┬─────┘
│ │ │ │
└──────────────┬───┴────────────────┴───────────────────┘
│ same MCP surface
▼
┌─────────────────────────┐ ┌────────────────┐
│ ashlr-plugin (10 MCPs) │ ←── │ LM Studio │
│ efficiency / sql / │ │ Qwen3-Coder │
│ bash / tree / http / │ │ -30B :1234 │
│ diff / logs / genome / │ └────────────────┘
│ orient / github │
└─────────────────────────┘ ┌────────────────┐
│ Ollama :11434 │
│ (fallback) │
└────────────────┘
All four agents talk to the same LLM and the same MCP tools, so behavior is consistent regardless of which one you launch.
Requirements
- macOS 14+ (Apple Silicon recommended; Intel works but slower)
- Docker Desktop — for OpenHands
- LM Studio with
qwen/qwen3-coder-30bloaded — primary LLM - Bun ≥ 1.1 — for ashlr-plugin MCP servers and ashlrcode
- Python 3.12+ — for Aider
- Node ≥ 20 / npm — for
npm install -g ashlrcode - ~32 GB free RAM (Qwen3-Coder-30B in 4-bit needs ~24 GB live)
- ~30 GB free disk (Docker images + model)
- Optional: Ollama as fallback LLM
- Optional:
ghCLI for GitHub PAT (GITHUB_TOKEN="$(gh auth token)")
Installation
# Plugin (provides the MCP servers all 4 agents share)
git clone https://github.com/ashlrai/ashlr-plugin ~/Desktop/ashlr-plugin
cd ~/Desktop/ashlr-plugin && bun install
# Workbench
git clone <this-repo-url> ~/Desktop/ashlr-workbench
cd ~/Desktop/ashlr-workbench
cp .env.example .env
# Per-agent installers (only run the ones you want)
./scripts/install-goose.sh # Goose via Homebrew
pipx install aider-chat # Aider (or: pip install --user aider-chat)
npm install -g ashlrcode # ashlrcode
# OpenHands needs no install — `aw start openhands` pulls the image on first run
After install, verify:
./bin/aw doctor # actionable diagnosis
./bin/aw health # full 13-point check
Usage examples
| Scenario | Best agent | Command |
|---|---|---|
| "Refactor this one file's error handling" | Aider | aw start aider . |
| "Add tests, run them, fix until green — autonomous" | OpenHands | aw start openhands then GUI prompt |
| "Quick interactive session, mostly tool calls" | Goose | aw start goose |
| "My usual driver — hooks, recall, fast" | ashlrcode | aw start ashlrcode |
| "Open a PR for me end-to-end" | OpenHands | give it a GitHub URL in the GUI |
| "Explain what this codebase does" | Goose | aw start goose → "orient on cwd" |
Configuration
Per-agent configs live under agents/:
agents/openhands/config.toml # OpenHands runtime settings
agents/openhands/mcp.json # MCP servers wired into OpenHands
agents/goose/config.yaml # Goose source-of-truth (copied to runtime on launch)
agents/aider/aider.conf.yml # Aider model + UX
agents/ashlrcode/settings.json # ashlrcode overlay (XAI primary, LM Studio fallback)
Workbench-wide environment lives in .env (see .env.example).
Troubleshooting
| Symptom | Fix |
|---|---|
aw start openhands says Docker not running | Launch Docker Desktop, wait for whale icon, retry |
| LM Studio "endpoint not responding" | Open LM Studio → Developer → Start Server, load the model |
| MCP server fails to start in any agent | cd ~/Desktop/ashlr-plugin && bun install |
| OpenHands GUI loads but agent errors on first turn | Confirm qwen/qwen3-coder-30b is the loaded model |
aider not found after install | pipx ensurepath or add ~/.local/bin to PATH |
ashlrcode complains about XAI_API_KEY | Set in .env or run with LM Studio fallback |
For anything else, run aw doctor — it prints the exact fix for each problem
it detects.
Contributing / extending
To add a new agent:
- Create
agents/<name>/with the agent's config file(s). - Create
scripts/start-<name>.shthat launches it pointed at the workbench config and the LM Studio endpoint. - (If it has a daemon) Create
scripts/stop-<name>.sh. - Add it to the
casestatements inbin/aw(require_agent,cmd_start,cmd_stop,cmd_status). - Add validation lines for its config in
scripts/healthcheck.sh. - Update this README's agent-lineup table and architecture diagram.
See CLAUDE.md for project-wide conventions.
License
MIT — see LICENSE.
Related Servers
Alpha Vantage MCP Server
sponsorAccess financial market data: realtime & historical stock, ETF, options, forex, crypto, commodities, fundamentals, technical indicators, & more
scan-mcp
Minimal MCP server for scanner capture (ADF/duplex/page-size), batching, and multipage assembly
Cloudflare MCP Server Example
A template for deploying a remote MCP server on Cloudflare Workers without authentication.
Kubernetes-MCP-Guard
AI-safe approval plan gated Kubernetes operations through MCP with OAuth, RBAC, audit, guardrails.
Kubernetes
Kubernetes MCP server with the top30 tools
Blockchain MCP Server
A server for blockchain interactions, offering Ethereum vanity address generation, 4byte lookup, ABI encoding, and multi-chain RPC calls.
Sequential Thinking Multi-Agent System (MAS)
An MCP agent that utilizes a Multi-Agent System (MAS) for sequential thinking and problem-solving.
Vibetest Use
A browser-agent QA swarm with an MCP interface for testing AI-generated websites.
UUIDv7 Generator
A server for generating version 7 universally unique identifiers (UUIDv7).
Node.js Sandbox MCP Server
Run arbitrary JavaScript in an isolated Docker container with on-the-fly npm dependency installation.
MCP Expert Server
An MCP server for query generation and documentation assistance using Claude AI.