Moneypenny is not another chat window. It watches your channels continuously, triages what matters, and hands signed-off work to whichever agent runner is plugged in this run — while the state, history, and decisions live one layer below all of them.
Five structural bets, each one a layer that only exists because the platform is always-on, cross-vendor, and built to outlast any one agent harness.
Every connected channel is watched continuously, without being asked. Signals get classified, entities extracted, and a ticket staged before anyone opens the thread.
The triage surface is not a dashboard to check. A five-second swipe cancels in-flight work, reprioritizes the queue, and feeds the decision straight back into the pipeline.
Claude Code, Codex, and pi are not just where work executes — they are where you touch the platform. Sign from a phone, close from a terminal: the mesh carries the same interaction everywhere.
A single skillset registers a watch-signal, an enricher, an execution verb, a datasource, and an outbound feedback hook at once — one plugin across five stages, not five separate integrations.
State and history live one layer below whichever agent is running. Every stronger model release becomes a dependency upgrade for this platform, not a competitor to it.
Ten axes, rebuilt on-brand from the internal scorecard. Moneypenny leads passive intelligence, human-in-the-loop triage, cross-vendor orchestration, multi-stage plugins, and subsumption resistance — and it concedes channel breadth, community, memory depth, and voice. Self-improving (measured substrate) sits low-mid: it built the measurement layer most systems skip, but hasn't closed the loop into autonomous improvement.
| Axis | Moneypenny | OpenClawa reactive gateway | Hermesa self-improving framework |
|---|---|---|---|
| Passive intelligenceLeads | 9/10 |
1/10 |
1/10 |
| HITL triageLeads | 9/10 |
2/10 |
1/10 |
| Cross-vendor orchestrationLeads | 8/10 |
4/10 |
2/10 |
| Multi-stage pluginsLeads | 9/10 |
3/10 |
2/10 |
| Subsumption resistanceLeads | 9/10 |
3/10 |
4/10 |
| Self-improvingmeasured substrateIn progress | 5/10 |
4/10 |
9/10 |
| Channel breadthConcedes | 3/10 |
9/10 |
5/10 |
| MemoryConcedes | 5/10 |
6/10 |
8/10 |
| VoiceConcedes | 3/10 |
7/10 |
3/10 |
| CommunityConcedes | 0/10 |
10/10 |
7/10 |
An internal scorecard, not a benchmark — these are our axes and our read of OpenClaw and Hermes, scored by us. The receipts behind every claim live on the guarantees page.
A skillset is not a single-stage plugin invoked on request. It registers across the pipeline at once — a watch-signal on intake, an enricher on triage, a verb on the execution engine, a datasource on entity extraction, and an outbound hook back to a channel.
The spine is purple everywhere on this site for exactly one reason: it is the color reserved for a skillset extension point, nowhere else. friendship-graph and media-finder cross the same five columns.
Three claims about how this platform holds itself accountable — each one backed by a mechanism, not a policy.
The mark, the wordmark, and the icon system this platform draws from.
Brand kit — packaged, versioned downloads land here when the release is cut.
Honest about what's not here yet: the brand kit above is a preview — packaged, versioned downloads land here when the release is cut. The scorecard is a rebuilt internal self-assessment, not a third-party benchmark — its rival columns describe categories of competitor, not a named product. The multi-stage skillset diagram illustrates career-partner's real hook points; friendship-graph and media-finder register fewer of the five today.