Changelog
What changed in NoClick: the SDK, the API and the app, newest first.
An agent fed by a trigger does not need a Message
Message is where you write an agent's standing instructions, and it is still required when nothing feeds the agent. When a trigger is wired into the agent, the event that arrives is the turn, so Message can be left empty and the node is no longer marked incomplete for it. Running such an agent by hand, with no event and no standing instructions, fails with an error saying so rather than running on nothing.
Improvements and fixes3
Several messages arriving at once no longer disturb an agent whose sandbox is still starting. A burst delivered while the agent was cold could stop the sandbox mid-start and lose the turns attached to it.
A turn lost because its sandbox was replaced is now recorded as a failed run with the reason, instead of sitting in history looking delivered.
A CLI agent with no account connected says which credential it needs. It used to fail with a message about model and tool limits, which pointed at the wrong thing.
Improvements and fixes1
The builder can add an Alarm or Filesystem node when you ask for one. Neither was in the list of nodes it could see, so asking it to let an agent schedule its own wake-ups used to get the answer that no such node exists.
Improvements and fixes2
Pressing Run on an integration node that is wired as a tool provider no longer fails with an error about a missing action. The button is off, and says the node provides tools and its actions run when the agent or your MCP client calls them. Run the agent instead. Asking NoClick to run that node from an MCP client gets the same explanation back.
Agents that run on a saved model credential work again. Since the previous release, an agent using an OpenRouter key or a Codex, Claude Code, or OpenCode sign-in failed at the start of every run with an error about its credential type.
Improvements and fixes2
An Instagram Login connection is treated as one everywhere it is used, including when an AI agent calls an Instagram operation as a tool. Its token could be sent to Facebook's API instead, which rejects it.
Sending an agent a second message while it is still working no longer leaves it stuck as working. When the agent takes several messages into one turn, the reply is matched to the messages it actually read, and Inputs consumed in the run's results lists them.
Improvements and fixes1
Both CLI dropdowns show plain model ids again. The Codex labels had picked up Codex's one-line model descriptions, which the picker printed a second time under each row and the closed field repeated as a sentence.
Improvements and fixes1
A Discord On Channel Message or On Bot Mention trigger scoped to one channel now fires for messages posted in the threads opened in that channel. Discord delivers a thread message under the thread's own id, so a reply inside a thread the agent had started was measured against the channel id and dropped. Leaving Channel empty was the only way to hear those replies.
Improvements and fixes2
The What came in section of a run's results shows the message the agent turn actually took in. A turn from a chat agent could show something else there: a note left behind by an earlier manual run, or a message from a different conversation. When several messages feed one turn, the section marks the one it shows as the latest of them, and Inputs consumed carries the rest.
An agent turn that never comes back shows in your run history as a failed run, with an error saying the turn was lost and that sending another message retries it. Those turns used to leave no run behind at all.
Improvements and fixes2
A Slack or WhatsApp agent that stopped reacting to messages works again. NoClick caps how many runs one channel or chat can start in a five minute window, so a reply loop with another bot cannot run away, and that counter could stop resetting. Once it did, every later message in that conversation was accepted and then dropped, with no run in Logs to show for it. The counter now rolls over as intended, and one that is already stuck clears itself on the next message.
The working indicator in an agent chat stops when the answer does. A finished turn could keep the agent showing as working under a reply that had already arrived in full, and it stayed that way until you reloaded the page.