agent mcp studio shipped on hn - multi-agent mcp systems running in a browser tab via pyodide + duckdb-wasm. solo founder build. the architecture is correct - browser-side execution removes the deploy step entirely.
what's elegant about it
no server, no auth dance, no deploy pipeline. a user opens a tab, configures agents, and ships work. for prototyping and personal use it's the cleanest mcp interface i've seen this year.
what becomes hard the moment a paying customer turns it on
- audit log - browser sandbox by default doesn't persist across reloads, the tamper-evidence story is ambiguous
- attribution - which agent did what, when, in whose tab
- compliance - did the agent touch pii, and can you prove it didn't
- task memory - cross-session persistence in indexeddb is fragile
the two integrations that close the gap
- mnemopay sdk for cross-session memory and reputation
- bizsuite ai-audit hook that emits a structured log to a hosted endpoint per agent invocation
await mnemopay.remember({ session_id, agent_id, action });
await bizsuite.audit.log({ session_id, agent_id, input, output });
why both ship together
the audit log proves what happened. the memory tells the next session about it. they're the same data, viewed by different consumers (auditor vs agent).
browser-tab mcp is the right idea. it just needs the audit and the memory before a fortune 500 will run it.
Top comments (1)
Some comments may only be visible to logged-in visitors. Sign in to view all comments.