Agent Activity: A User Activity Recognition Toolkit for Claude Agents

Released this week on GitHub and picked up on Hacker News, Agent Activity is an open source toolkit developed by Rafa Gómez Guillén that addresses a specific challenge in agent design: giving agents context about what the user is doing in their work environment, not just what the user explicitly tells the agent.

The premise is straightforward. A Claude agent that only receives text instructions operates with incomplete information. If the agent knows the user has been editing a configuration file for ten minutes, just switched directories, or executed a series of related commands, it can respond more accurately and anticipate the next need better. Agent Activity aims to bridge that gap.

What it does exactly

The toolkit functions as a layer for collecting signals of user activity, designed to integrate with agents running on Claude Code or as an MCP server. According to the repository, it captures events such as:

• File activity (opening, editing, saving)
• Command terminal history
• Work context changes (active directory, project)
• Session metadata

This information is structured and exposed as context consumable by the agent, so it can incorporate it into its reasoning without requiring the user to verbalize it in each prompt.

The name "User Activity Recon Toolkit" is deliberate: it borrows vocabulary from operational intelligence to describe what it does, which is essentially observing and synthesizing environmental signals before the agent acts.

Why it matters for the Claude ecosystem

In the current workflow model with Claude Code, hooks and MCP servers are the natural mechanisms for enriching the context an agent receives. A `PreToolUse` hook can, for example, inject system state information before the agent calls a tool. Agent Activity appears to operate at that layer: it does not modify model behavior, but rather improves the quality of input context.

This connects with a trend we've been observing in the ecosystem for months: teams that get the best results from their agents are not necessarily those using the largest model, but those building richer and more precise context pipelines. With Claude Opus 4.7 and its million-token window, the ability to absorb context is no longer the bottleneck; the bottleneck is what context is worth including.

Who this makes sense for

The most immediate use case is developers building technical support or local workflow automation agents. If your agent needs to understand where the user is in their work to decide what to do next, a toolkit like this can save considerable ad hoc logic.

There is also an interesting angle for teams building specialized sub-agents within Claude Code: a code review sub-agent, for example, could benefit from knowing which files the user has touched in the minutes before receiving the formal task.

It should be noted, however, that the repository has just been published and traction on Hacker News remains minimal (one point, no comments at the time of collection). There is no extensive documentation or proven production use cases to cite. It is a very early stage project.

Privacy considerations

Monitoring user activity locally to feed an agent is useful, but it also involves design decisions that the repository should explain clearly: what gets recorded, for how long, whether data leaves the local environment, and at what granularity. In corporate environments or on shared machines, these aspects are not optional.

The project deserves monitoring if the author maintains development pace and adds documentation on integration with Claude Code and MCP. The idea is solid; execution remains to be seen.

---

Original source: Agent Activity on GitHub, picked up on Hacker News on May 26, 2026.

EP Opinion: The approach of enriching agent context with real user behavior signals is one of the most practical problems in agent development today. Having someone address it with a dedicated open source tool is a positive signal, though the project needs to mature considerably before becoming an ecosystem reference.