Open Source

Some Useful Agents

A local-first agent operating system for building, running, and supervising useful AI workflows.

Pulse dashboard showing runs today, failure rate, agents, and scheduled work — Pulse — runs, failures, and scheduled work at a glance

Agent overview page with the flow graph of nodes — Agents are flows — a graph of shell, LLM, and tool nodes

Run detail page with per-node execution output — Every run is recorded — timing, trigger, result, and which nodes ran

1 / 3

The problem

Most agent tooling is good at producing one answer and bad at everything after that. Runs disappear into terminal scrollback, prompts live in someone's clipboard, and when something breaks there's nothing to inspect.

The hard part of agent work is not generating one answer. It is building systems people can rerun, supervise, debug, and extend.

How it works

Some Useful Agents turns ad hoc agent work into reproducible, composable workflows — and makes the operational layer explicit.

Agents as flows — each agent is a graph of nodes that run shell commands, LLM prompts, tools, or other agents
Run it anywhere — from the CLI, from the dashboard, or over MCP from Claude Desktop and other clients
Schedule it — cron on any agent, with pause and resume so nothing runs invisibly
Supervise it — the Pulse dashboard for runs, failures, and replays

What's in the box

Multi-provider fallback — Claude, Codex, on-device Apple models, or any OpenAI-compatible endpoint
Local-first secrets (encrypted at rest) and global variables
MCP servers as first-class tools — paste a config, pick the tools
Reusable packs of agents and dashboards, plus 40+ example agents

Interfaces CLI, dashboard, MCP

Runs Locally, on your machine

License MIT

Source GitHub

Try it with npx @some-useful-agents/cli@latest init or clone the repo.

Read the case study for the full backstory.