The local version & MCP
Run StepBerry on your own machine, bring your own AI, and connect over MCP.
Run StepBerry on your machine
The local version is the same app on your own computer: procedures live in a local database, files on your own disk, and nothing is sent anywhere. It runs single-user by default — a password prompt on your own machine guards nothing — and can turn real sign-in on for a shared installation.
The local version is in early access alongside the rest of the paid tiers — it ships as a small Node app. The about page has the honest state of what's built.
Connect your own AI
Locally, the assistant speaks to any OpenAI-compatible endpoint. Point it at Ollama, LM Studio, or anything else that answers on the same protocol:
STEPBERRY_AI_URL=http://localhost:11434 # default: Ollama
STEPBERRY_AI_MODEL=llama3.2The cloud version includes a hosted model; no configuration needed there.
MCP: let your tools talk to your manual
StepBerry is an MCP server. A local AI client (Claude Desktop, an editor agent, anything MCP-aware) can search procedures, read sections, and manage canned answers through it. Locally it runs over stdio:
{
"mcpServers": {
"stepberry": { "command": "node", "args": ["src/mcp.ts"] }
}
}MCP over HTTP
The cloud workspace exposes the same tools at POST /mcp, guarded
by an access token. Create one under Settings → Access tokens and send
it as a bearer:
Authorization: Bearer sbk_…A token carries the permissions of the account that made it — the tool list itself is filtered, not just the calls. No token, no answer: requests without one get a 401.