The idea came from trying to get agents (e.g., OpenClaw, Hermes Agent) to perform more useful tasks than simply organizing our calendars or monitoring our emails. We wanted them to answer questions and take actions that depend on timely external data, e.g., “Can you find me roundtrip flight options from NYC to SF next month”.
Personal assistant agents appear to be great frameworks for maintaining user preferences and acting with that information (e.g., “Justin hates early morning flights,” “Justin doesn’t mind a layover if it’s shorter than 3 hours,” etc), but existing tools for providing assistants with actionable, external data are surprisingly limited.
For example, browser automation can work, but it’s generally slow, expensive, and flaky [1].
Ad-hoc scraping works well in narrow cases, but can become quite expensive at scale.
MCP servers can inject tool schemas into context on every turn (inflating token consumption), and are often either too granular or too coarse in their design.
So we built agent-data around a simpler interface: give the agent the ability to discover the right endpoint at runtime, call it via CLI, and get back compact structured output.
For example, to answer my earlier question, an agent might use: `agent-data call search-flights --origin_iata JFK --destination_iata SFO [--params]`
We chose the CLI design because agents already work well with bash. Further, commands can be composed, piped to files, and are generally more flexible than off-the-shelf MCPs. Instead of teaching an agent to use a browser or learn a new MCP tool schema, we wanted to give it a smaller, more predictable interface.
The current version supports: - Live data endpoints for categories like flights, restaurants, RSS/blog feeds, and social media feeds - Structured JSON outputs by default - Endpoint discovery so agents can “find” relevant capabilities instead of loading huge lists of tools into context by default - A CLI interface designed for agent workflows
It’s still early and some parts are a bit rough, but we’re adding new endpoints and expanding the scope of existing ones every week.
We’d love to know: What live data sources do you wish your agents could access? Where have browser automation, scraping, generic web search, or MCP tools failed in your own agent workflows?
No comments yet.
