August 2026
Everything that changed in NoClick in August 2026.
Improvements and fixes2
Connecting a ChatGPT account to a Codex agent now checks that the sign-in carries the identity token Codex needs to reach ChatGPT. OpenAI does not always return one, and without it the connection saved but the agent ran against the OpenAI API instead, so its turns failed with an authorization error. NoClick fetches the missing token before finishing, and asks you to sign in again if it cannot.
A ChatGPT account you connected earlier repairs itself. When the identity token on an existing connection is missing or has expired, NoClick refreshes it on the agent's next run, so a Codex agent that stopped working starts again without you reconnecting it.
Improvements and fixes1
An external MCP server that answers over a held-open stream now reaches your agent. Servers of that shape used to leave the agent with none of their tools for the whole session, and could leave it stuck rather than answering.
Improvements and fixes2
The built-in LLM agent has the
upload_filetool at all times. It used to appear only when a filesystem node or a GitHub repository mount was wired, so an agent could write a file in its sandbox and then have no way to turn it into a shareable URL.A shell command with very large output no longer breaks a conversation. One oversized result could grow past the model's context limit and get saved into the conversation, after which every later message failed, including a plain "hi". Command output is now trimmed before the agent sees it, keeping the beginning and the end, and conversations already affected recover on their own.
Improvements and fixes1
An HTTP Request node wired into an agent as a tool can send a body again. Its POST, PUT, PATCH, and DELETE tools now take the body, form fields, and content type, so the agent's requests are no longer sent empty.
Improvements and fixes1
The Test Run screen no longer spins forever when an agent replies with structured output, and it starts the test run you asked for instead of the first one on the workflow.
The key an agent runs on matches the model you picked
An agent node can hold connected keys for more than one provider, and switching the Model could leave an earlier key attached. That key was used anyway, so the run failed on a rejected or missing API key even when the right key was connected.
NoClick now runs the agent on the credential belonging to the provider of the model you picked. A key for any other provider is ignored, and the run behaves as if nothing were attached: models NoClick serves on its own keys keep running on credits, and models that need your own key stop with a missing credential you can fix in the node's Credentials tab.
Improvements and fixes1
Agents on Gemini 3 models can use tools again. Every turn where the agent called a tool failed with an argument error from Google, which left agents answering channel messages silent.
Improvements and fixes1
An agent the AI builder runs to test it keeps the tools wired into it. The run dropped the connection between the agent and the services feeding it, so the agent executed with none of the actions you had allowed and reported that it had no tools.
Improvements and fixes2
Asking the builder for an agent on a specific model now sets that model. Naming Claude used to give you an agent called "Claude" running the default model instead. Agents are also no longer named after a model vendor, since that name goes stale as soon as you switch models.
Triggers from PostHog, Sentry, Honeycomb, Calendly, Klaviyo, and BambooHR now reach an agent they are wired into. The run used to fail the moment the trigger fired.
A stalled agent turn no longer bills for hours
A CLI coding agent's sandbox is billed for as long as it stays up, so a turn that started and never delivered a reply used to keep costing 0.05 credits per minute until the four-hour sandbox limit. Those turns now end: the sandbox shuts down as soon as the agent's process is gone, and a turn that produces no output at all for 30 minutes is treated as stalled and stopped.
- The chat reports the turn as lost and unlocks, so you can send the message again, instead of showing a reply that is still coming and never arrives.
- Files the agent saved to its workspace survive the stop, as they do for any other sandbox shutdown.
- A turn that is genuinely quiet for 30 minutes, for example one long build inside a single command that prints nothing, is stopped as well and has to be sent again.
Improvements and fixes1
An agent's tools now keep working for the whole conversation. In a long session, integration tools, MCP servers, and NoClick's own agent tools used to stop responding after about an hour and report that they needed re-authorization, which left the agent explaining what it could not do rather than doing it.