Hello everyone.
A proxy is easy to use until something fails. Then the dashboard becomes a wall of numbers: a connection that took 1.2 seconds, a DNS lookup that took 800 milliseconds, a rule that matched, a policy that answered, an outcome with a name. The numbers are honest. The reason is not in them.
Chute 1.5.0 adds Apple Intelligence, running on the phone. It explains a diagnosis, explains a connection, looks for a pattern in recent connections, proposes fixes for configuration errors, and drafts rules from a sentence. Nothing it reads leaves the device, and nothing it writes is trusted without being checked.
Where It Shows Up
Five places, each with a way to ask Apple Intelligence, and a different job:
- Diagnosis Summary. After a network or proxy diagnosis, it reads the rows and says whether the network is healthy, and if not, what looks wrong. It is told to mention only failures and clearly abnormal values, and to give at most four findings with something to try. (If you have not used the diagnosis pages, this post explains what they test.)
- About This Connection. One connection from the dashboard: which rule matched, which policy carried it, how it ended. If the outcome is a failure or the speed is slow, it adds the likely cause and what to try; a connection that ended normally, or was blocked on purpose, gets no invented advice.
- Recent Connections. A summary of what went wrong lately: a host or policy that failed more than once, a policy that is often slow to connect, slow DNS. One failure is not a pattern, and the prompt says so.
- Configuration errors. When the editor rejects a file, it explains each error and proposes a fix: replace the line, delete it, or mark it as needing a manual edit.
- Rules. “Describe in Words…” turns a sentence into rules. The placeholder suggests: “send Netflix through my US group, and block doubleclick.net”.
A Connection, Explained
Say a video app will not load. The dashboard shows the connection: a rule matched, a policy carried it, the request went out, nothing came back before the connection timed out, and the node took 1.4 seconds to answer.
Tap the button, and the answer is a short paragraph and a few lines: the rule and the policy that carried the connection, the fact that the upstream never answered, the 1.4 seconds as slow for this network, and two things to try — another node in the group, or a check of whether the site is reachable from that region. No addresses are invented, and nothing is offered as a certainty the data does not support. If the data does not show a cause, the answer says that too.
That shape — a verdict, then findings, then something to do — is the same on all five screens. It is deliberately short. A proxy’s dashboard is not the place for an essay.
The Parser Has the Last Word
The configuration features are the ones that could do damage, so they are the ones with the least freedom.
A fix is not text the model writes into the file. It comes back as a structured proposal for a specific line: what to do with it, and what the replacement is. The app then applies the proposal to the file and parses the whole thing again with the same parser the tunnel runs.
A fix is marked verified only when the edited line’s error is gone and no other line has gained one. A deletion is verified when the error it was meant to clear is gone. If the parse still fails, the parser’s own message goes back to the model once, so it can try again with the reason; after that the fix is shown as “Not verified” or “Needs a manual edit”. The sheet counts them: “3 of 5 fixes verified by Chute’s parser.” Nothing is applied until you tap Apply, and the text stays editable afterwards.
Rules work the same way, with one more step. The model sees your sentence, the names of your policies and groups, and up to 25 hosts the phone actually reached recently. Each draft is normalised — a domain, a CIDR range, a GEOIP entry — and re-parsed. The sheet says what a rule would have matched: “Matches 12 recent connections to 3 hosts.” You add them, or you do not.
The schema is part of the fence. The set of policies a rule may name is the configuration’s own policies plus DIRECT and REJECT, so a rule pointing at a policy that does not exist cannot come back at all.
Why the App Does the Counting
The first version taught us where the model should not be trusted.
It was handed the dashboard’s connections with their close reasons translated by name rather than by meaning. A normal upstream close — how most connections end — went in as “FAILED: closed by the proxy server”. A timeout, mostly the fifteen-minute idle limit, always counted as a failure. A connection Chute closed because you switched nodes read “closed by the user”. On a sample of 43 connections with 5 real failures, the summary said 26 had failed, and the model duly blamed the node.
Phases that were never reached had no timestamp, and subtracting “not reached” from a time gave the time since the tunnel started: in global mode, or after a failed lookup, every connection read “rule 3806 s”. One string described Chute’s own local listener as the proxy server. And the prompt asked the model to find patterns in about a hundred raw rows — something a small on-device model cannot count.
So judgement and counting moved out of the model. A digest step, outside the model and tested without one, decides whether a connection failed or was slow, counts failures by cause, policy and host, and measures connect and DNS time by policy. The model receives those conclusions, plus the failed and slow connections with the engine’s own notice. Connections that worked are not listed. Both prompts now say: keep to Chute’s judgement, do not recount, one failure is not a pattern, and if nothing is wrong, say so in one sentence.
What It Will Not Do
It will not rewrite your configuration in the background. It will not add a rule you did not see. It will not invent a server address, a password or a URL — the prompt forbids it, and anything that becomes configuration is checked before it is shown.
It is also not a general chatbot. There is no conversation to have and no follow-up question; each button is a single request with the data on that screen. That is a deliberate limit. A proxy’s job is to move traffic, and an assistant that can act on the proxy is a much larger surface than one that explains it.
On the Device
Everything runs on Apple Intelligence’s on-device model. We do not use Private Cloud Compute, and no diagnosis, connection record or configuration text is sent anywhere. The app says so under each answer: “Written on this iPhone by Apple Intelligence; nothing was sent anywhere. It can be wrong, so check before you rely on it.”
It needs iOS 26, a device that supports Apple Intelligence, and Apple Intelligence turned on. If any of those is missing, the buttons are not shown, or the reason is explained: the device is not eligible, Apple Intelligence is off in Settings, or the model is still getting ready. Turning Apple Intelligence on does not restart Chute; the buttons appear the next time a screen is shown.
The app’s eight languages are supported where the model writes them; where it does not, it answers in English rather than in a broken translation. Long inputs are trimmed with a note saying how many lines were left out, and if the model says the input is too long, the app retries once with half of it. Apple’s guardrails can decline a request; when they do, the app says so instead of showing an empty box.
What This Is For
The dashboard tells you what happened. The new part is a sentence about why, in the place where you were already looking, and a fix that has been through the same parser that will run it.
Chute 1.5.0 brings this to iOS 26 devices with Apple Intelligence on.
Thanks.
Chute Devs