Changelog
What is new.
Every release of TrailSail, newest first. The app shows the same notes when it offers you an update. How updates work.
One short email for each release, and nothing else. You confirm by email first, the list is kept with Mailchimp, every email has an unsubscribe link, and your address is never passed on. Protected by Google reCAPTCHA, which loads when you start filling the form in. Google's Privacy Policy and Terms of Service apply. Privacy
Less waiting before a reply starts, on Splash. The server package stays at 0.4.1.
New
- Reading ahead after a reply. A model server skips text it has already read, but TrailSail changes earlier text between turns on purpose: older tool results are cut down once they are a couple of turns old, and a long chat is folded into a summary. Until now the server read everything after such a change again at the start of your next message, while you waited. Now, when a reply ends and something earlier is due to change, TrailSail sends the next request's opening to be read while you are reading the reply. Your next message then starts almost at once.
- It is sent once per reply and never repeated, however long the chat sits idle.
- It is not sent while anything is being written, because reading slows writing.
- It is skipped when the server has nothing worth reading.
- Splash only. Turn it off in Settings → Context → Read ahead after a reply.
Changed
- Agents share what they have in common. Every sub-agent is sent the same instructions and the same list of tools, and used to be sent its own task before them, so each agent read all of it again. The task now comes after the shared part, and agents started together read that part once between them.
Measured on a 32 GB Mac mini with a 27B model: 12,000 tokens took 36 seconds to read cold and under half a second once read. Four requests that open with the same 12,000 tokens finished together in 31 seconds, the time one takes alone; four that open differently took 114 seconds.
Web search and page fetching that hold up when several things are searching at once. The server package stays at 0.4.1.
Fixed
- Searches no longer arrive in bursts. Sub-agents working side by side each ran their own searches at the same moment, which is what gets a search service to refuse or suspend you. Every search now goes through one queue: one at a time, with a short pause between them.
- The same search is not run twice. A question asked again within ten minutes is answered from memory, and the same question asked by two agents at once is asked once.
- A busy search service is tried again. When it says "too many requests" or does not answer, TrailSail waits a few seconds and tries twice more before giving up.
- A backup search provider. If you have a Brave or Tavily key saved as well as SearXNG, searches move to it whenever the chosen provider does not answer, and go back when it recovers. Settings → Web search explains how to save one.
- SearXNG with its engines suspended is reported as that. It used to come back as "No results", as if nothing had been found.
- The model is told not to scrape search engines. When search failed, the model would try to read a search engine's results page directly, which those sites refuse and count against your connection. That is now stopped before any request is made, and the model is told to use web search instead.
- Pages are fetched one at a time from any one site, a second apart, and a site that says "too many requests" is asked once more after the wait it names.
- Clearer reasons when a page is refused. Each kind of refusal now tells the model what to do next, such as not guessing addresses after a "not found".
- The app names itself correctly. Page requests carried a link to an address that no longer exists; they now point to trailsail.app.
A small release for the app. The server package stays at 0.4.1.
Fixed
- Memory for models says why it is locked. After updating a server from TrailSail, the setting can still be locked, and the note read as if the installer had to be run on the Mac you were using. It now says plainly that one more step is needed on the server itself: open the trailSailServer installer there once and enter its password. The Server panel shows the lock too, without opening the dialog.
- Spacing in the Server panel. The licence line sat tight against the line above it, and the Memory for models row left a gap beneath it.
Changed
- macOS 14 or later. The app now says what the website has always said. Earlier versions claimed macOS 12.
This release fixes installing the server on a Mac that has never had Splash on it. After updating the app, open Settings → Server and press Update to 0.4.1.
Fixed
-
The server installs on a new Mac. The installer asked Homebrew for a package that does not exist, so on a Mac without Splash it stopped part-way. It now installs Splash from its makers, Inco AI.
-
Older Macs get the llama.cpp router. Splash needs an Apple M3 or newer and macOS 26.4 or later. On an older Mac the installer now says so, leaves Splash out and installs the llama.cpp router alone, where before it failed.
-
Splash always runs with its key. Splash serves without a key unless it is started with one. The server now starts it with the key, so only paired devices can use it. On a server that is already running, this takes effect the next time Splash is restarted from the Server panel.
-
No false "too low" warning. Memory for models said the recommended amount was too low for the model, because it overestimated what a model takes. The estimate is now based on what a model really uses, and the recommended amount is never called too low.
-
The memory slider always moves. On a server that cannot apply a change yet, the slider was locked. It now moves, so you can see what each amount would mean, and the dialog says plainly that the installer has to be run once on the server before a change takes effect.
Also
- TrailSail is now built for Apple silicon alone. It has always needed an Apple silicon Mac; the app no longer carries an unused Intel half.
- The guides now show how to run everything on one Mac, with Splash or Ollama, and the Splash steps in the manual setup guide are corrected.
You can now choose how much of the server's memory the models may use. This needs the server package 0.4.0, and the installer has to be run once on the server itself: download trailSailServer-0.4.0.pkg from trailsail.app and open it on that Mac. Updating the server from Settings → Server brings the new version but cannot install the part that changes the setting; the app tells you when that is the case. Engines, keys and pairings are kept.
New
- Memory for models. Settings → Server shows how much memory the models may use, with Change… next to it. A slider sets it in whole gigabytes, within a range worked out from the server's own memory, and Use recommended puts it back.
- Warnings when the figure is wrong for the Mac. The app says so when the figure is too low for the model the server runs, or so high that too little is left for macOS. If the server is the same Mac you are working on, it recommends less and warns sooner, because the models share memory with TrailSail and your other apps.
- The same on the server.
trailSailServer memoryshows the figure, the recommendation and the range, andtrailSailServer memory set 24GB(orauto) changes it. The server's menu bar item has Memory for models too.
Fixed
- The memory limit now fits the Mac. The installer used to set 27 GB on every Mac, which is more than a 16 or 24 GB Mac has and less than a larger Mac allows by itself. It now sets a figure for the Mac it is installed on, and leaves a Mac with less than 16 GB to macOS. A 32 GB server keeps 27 GB.
- The limit survives a restart. It is applied again every time the server Mac starts.
After updating the app, open Settings → Server and press Update to 0.3.1 to bring your server up to date.
Changed
- The icon is the logo on dark. While TrailSail is open and its appearance is light (Settings → General: Light, or System on a Mac in light mode), the Dock shows the logo on white instead.
Fixed
- No false warning in the server's menu bar. The llama.cpp router is only started when its model is wanted, so after a restart the server's menu bar item showed a warning triangle with nothing wrong. A router that has not been started now shows as "not started" and the sailboat stays.
- A new icon. TrailSail has a new logo. On macOS 26 the icon follows your appearance setting: on white in light mode, on dark in dark mode.
Nothing else changes in this version, and the server package stays at 0.3.0.
This release is about the server package. After updating the app, open Settings → Server and press Update to 0.3.0 to bring your server up to date. Engines, keys and pairings are kept.
New
- A menu bar item on the server. The Mac that runs the models now shows one icon in its own menu bar: a sailboat when everything is running, a warning triangle when something needs attention, a stop symbol when nothing is running. The menu lists each engine with its model, lets you start, stop and restart it, open its log, and make a new pairing code.
Fixed
- Installing on a new Tailscale account. The installer now checks that HTTPS is turned on for your Tailscale network before it installs anything, and tells you where to switch it on (Tailscale admin console → DNS → Enable HTTPS). Before, the install finished and showed a pairing code for a server that TrailSail could not reach.
- A server that is not published says so. If Tailscale could not publish an engine, the installer stops and names the reason, and
trailSailServer statusshows what is not published.
Also
./install.sh --checkruns the installer's checks on their own and installs nothing.
The first release published on trailsail.app.
New
- Signed and notarised by Apple. The app and the server installer open without a warning from macOS.
- Updates. TrailSail asks on first launch whether it may check trailsail.app once a day for a new version, and installs it when you agree. The server package is updated from Settings → Server.
- Licences. TrailSail is free for one person. A company with a team adds its Business licence in Settings → General → Licence.