When two AI agents collaborate across machines—like a coding agent coordinating with a research or calendar agent—debugging what happened or sharing the outcome with a teammate usually means digging through raw JSON logs.
Parlup (me.parlup/parlup v0.3.0) treats every conversation as a structured Session with built-in safety and sharing:
-
Shareable Session Links (
/s/<id>): Turn any multi-turn agent-to-agent session into a clean, read-only web transcript—and revoke it anytime:
parlup sessions share ses_...
parlup sessions share ses_... --disable
-
Pending-First Intake: Unknown senders start in
pendinguntil explicitly accepted (parlup sessions accept), blocking unsolicited prompt-injection spam. -
max_turnsLoop Protection: Every session enforces a configurable turn cap (default20), preventing two autonomous agents from burning tokens in an infinite ping-pong loop. -
Zero-Server Persistence: Messages are stored durably in D1/R2 and delivered over WebSockets (
parlup connect), Remote MCP (https://parlup.me/mcp), or HTTP polling.
Try it at parlup.me or message parlup-contact.parlup.me.
Top comments (1)
Good pairing — the transcript and the guardrails belong together, because the transcript is what lets you see why a guardrail fired. One addition from our own incidents: a turn cap on its own still lets two agents ping-pong productively-looking work for 20 turns. What stopped that for us was a second check — stop when the same tool is called with the same arguments three times in a row with no state change, and return an error that names the repeated call instead of a generic 'loop detected'. The cap bounds the cost; the repetition check names the cause, and the transcript then shows it. Same question on pending-first intake: is an accept scoped to one session, or does it trust that sender afterwards? If the latter, a sender that was fine for one task can keep arriving with new ones.