stairway

Most of my side projects start in a chat. I talk an idea through with Claude, usually on my phone, and at some point the conversation is good enough that I want to try it. That’s where the momentum used to die: get to a laptop, SSH into the server, make a folder, start an agent, paste in a summary of everything we just said.

stairway is the step between those two places. It’s a small MCP server running on my machine. When the chat says “ok, let’s try it”, Claude calls it, and a few seconds later there’s a folder, a git repository and a coding agent already working on the brief we agreed on.

The route

1. In the chat, I ask Claude to create an experiment and brief it. That’s one tool call: create_experiment(name, prompt).
2. On the server, stairway makes <root>/<name>, runs git init and starts the agent inside a detached screen, so it survives restarts.
3. Wherever I am, I follow it: in claude.ai/code on the phone, in a VS Code that stairway opens in the browser, or in a Telegram group where the agent’s answers arrive next to the people discussing them.
4. Later, the chat can keep steering it with send_prompt, and the status page shows every experiment, its terminal and what it was told.

What it is careful about

stairway exists to let an agent run commands on my machine when asked, so it’s built around explicit consent. Existing folders and Telegram groups do nothing until I approve them on the status page. claude.ai connects through stairway’s own OAuth server, whose grants are signed rather than stored, so one secret revokes everything. Agents never see stairway’s own credentials.

Not only Claude

Claude Code is the default. When I want a different model, or a cheaper one, an experiment can run opencode or crush instead, through OpenRouter or any OpenAI-compatible provider. Their replies reach Telegram the same way, because stairway reads each agent’s own conversation store rather than scraping its screen.

Where it came from

I built it in a few sessions of pair-programming with Claude Code, on a fairly old Ubuntu server. That explains some of its quirks: a byte-pipe proxy for VS Code, a little color filter because GNU screen 4 can’t show 24-bit color. It’s deliberately small: one file you can read in an evening, and no dependencies to keep up to date.