Field notes
What Smithery and ClawHub Usage Taught Us About Agent-Native Software
Agent software can receive meaningful machine-driven use while showing weak human social proof. Installs and calls often happen through automated paths that bypass repository pages, review forms, bookmarks, and star buttons. Measure usage, repeat integrations, defects, and attributable outcomes separately from public engagement.
Answer first
Agent software can receive meaningful machine-driven use while showing weak human social proof. Installs and calls often happen through automated paths that bypass repository pages, review forms, bookmarks, and star buttons. Measure usage, repeat integrations, defects, and attributable outcomes separately from public engagement.
The public numbers looked contradictory
As of August 3, 2026, the Sentien Labs ClawHub publisher showed approximately 3,800 lifetime downloads across four skills. The VerdictSwarm skill showed 928 lifetime downloads and 141 in its displayed 30-day window. Yet the publisher had no meaningful bookmark activity and no conventional review trail.
On the same date, Smithery showed approximately 213 reported VerdictSwarm tool calls. Its dashboard exposed operational aggregates, but not a conventional product-review surface. These numbers are a dated snapshot and will change after publication.
The apparent mismatch was not evidence that usage was fake or that users disliked the software. It showed that agent software has two different participants: the machine that installs or calls the tool and the human who owns the social account.
Machine activity is not the same as human adoption
A tool call confirms that a client reached a named capability. A session can include repeated schema discovery, availability testing, automated probes, or a real workflow. A download can come from a fresh install, CI, a cache miss, or a human operator. None of those events alone proves a retained customer.
Smithery also identified listability testers and schema-oriented clients among its activity. That makes the totals useful for detecting demand direction and operational health, but insufficient for counting people or revenue.
Why reviews and stars lag
- The agent can invoke a tool without opening its marketplace page.
- Install commands and client catalogs bypass the repository README.
- The human may see only the final answer, not the source that produced it.
- ClawHub emphasizes bookmarks rather than a familiar five-star review flow.
- Smithery's public surface emphasizes connection, health, and issue reporting rather than reviews.
Asking every invocation for a review would make the tool worse. Social requests belong after demonstrated value and in human-facing documentation, not in the critical path of an automated risk decision.
The metrics need separate columns
Reach
Downloads, package installs, listing views, and unique discovery sources.
Use
Tool calls, successful responses, repeat sessions, latency, and failure rate.
Human engagement
Bookmarks, stars, issue reports, discussions, and integration posts.
Outcome
Whether the tool changed a decision, prevented an error, or earned retained usage.
Mixing these columns creates bad decisions. A spike in probes is not product-market fit; zero reviews are not zero usage; a star is not evidence that the tool changed an outcome.
What we changed
- Published a public source repository for the exact OpenClaw skill bundles.
- Aligned Smithery and MCP discovery with the current three-tool contract.
- Kept migration responses for retired MCP tool names instead of silently breaking older clients.
- Added first-party pages that explain which integration path a human operator is choosing.
- Prioritized integration stories and defect reports over generic review requests.
A better feedback request
The most useful question is not “Will you leave five stars?” It is: “What did your agent use, what happened, and did the result change the next step?” That can uncover a false positive, a missing field, a broken client, an adoption story, or a reason the tool did not earn repeated use.
If the tool has been useful, a bookmark or repository star helps discovery. But honest integration evidence is the stronger proof—and it should never be purchased, faked, or made conditional on positive sentiment.