What you need

The apps

You need all three apps installed through Scalpels, into the infrastructure that runs the software you want to improve, and connected to it:

• Hone receives your app's telemetry. See getting started with Hone.
• Wake receives your app's logs. See connecting your apps to Wake.
• Reel records browser sessions. See connecting your apps to Reel.

You can start with any one of the three. The loop gets better with each app you add, because each one answers a question the others cannot. Hone tells you a request got slow. Wake tells you what the app was saying when it happened. Reel shows you what the person on the other end was doing.

A read-only Scalpels token

Create a token on Agent access (/settings/agent-access) and give it Read installed apps, not Call installed apps.

This matters more than it looks. A read token can call any read tool on any app your team has installed, and nothing else. Wake has tools that delete logs and rotate credentials. A read token cannot reach them, no matter what ends up in the agent's context. Nothing in this loop needs more than read, so do not give it more than read.

Install the token in each agent runtime that Solo launches, the way Use Scalpels from your agent describes. Then check it by hand from a Solo agent before you automate anything:

Using the Scalpels MCP, list my teams, then the installed apps for my team, and tell me which ones are relay ready.

If Hone, Reel and Wake all come back relay_ready, the conductor will be able to reach them.

Solo, on a Mac that stays awake

The conductor and the investigators are Solo agents, so Solo has to be running whenever the loop does. Put it on a machine that stays on: a Mac mini in a cupboard is ideal. Turn off sleep, sign in so Solo starts at login, and give it a project for the repositories you want improved.

Solo also ships an MCP server as a binary inside the app, at /Applications/Solo.app/Contents/MacOS/mcp. That is how the script on cron reaches Solo from outside it: it starts the binary, speaks MCP to it, and uses the same tools your agents use to check on the conductor, start it and talk to it.

Ask an agent in Solo to call list_agent_tools. It lists the agent runtimes Solo can start, by name. The conductor's configuration names one of them, and the conductor uses the same list when it starts investigators.

The companion code

Clone the continuously-improve-your-software branch of artisan-build/scalpels-guides onto the same Mac. It holds the script cron runs and the files of instructions the agents read, and it needs nothing beyond the Python that ships with macOS. The conductor walks through setting it up.

The GitHub CLI

The conductor files issues and the investigators open pull requests, both with gh. Sign it in on the Mac as an account that can open issues on every repository involved, and nothing more. If you would rather the automation had its own identity, a machine user with access to those repositories is cleaner than your own account.

A map of your apps

The same software has a different id in each tool:

Write these down once, in the ledger, along with the GitHub repository each app's code lives in. The conductor reads the map on every run, and investigators use it to know which repository to work in.