Ancilo 0.6.0

• Security maintenance: updated TLS and ZIP dependencies to versions that fix newly reported vulnerabilities; updated the spreadsheet reader and writer to use the repaired ZIP library.

• Chat answer quality: ordinary writing, explanations and plans no longer enter system controls because of fragments such as “hot” in “photo” or inherit tools after changing the topic. Local reasoning models get a bounded thinking allowance to check their answers; this can increase response time. Writing and follow-up instructions now emphasize preserving facts and constraints.

• Everyday chat: unsent drafts survive switching views and reloading the same window. Failed sends preserve newly typed follow-ups; a completed answer no longer pulls you back from another view. Filter conversations by title (including chats in folders), and copy a whole answer with one click.

• Checks distinguish results within one task: instructions naming a spreadsheet no longer require its columns in a separate Word report. Shared instructions and later corrections retain their file scope in both previews and the combined check. Textual “Gesamtkosten” totals are recognized without claiming their arithmetic was verified.

• Interrupted tasks can continue in their existing copy: accepted requests and file references are saved before the model starts. Lost replies no longer duplicate task creation or messages when retried from the Tasks view. Partial uploads can be retried, including after reloading the window; missing files are requested again.

• Task recovery is visible: interrupted tasks offer Continue task, the task view stays available after a refresh error, and stale live progress is cleared when the event connection drops.

• Document tasks receive clearer path instructions: filenames without a directory are written at the top of the working folder. Document tasks no longer receive an unrelated absolute project path; Word text is instructed to use real paragraph breaks.

• Result checks recover after an error: a failed check after files changed no longer sends repeated requests. Tasks offer Check again, hide outdated check badges, and enable Keep or Save only when the current results have been checked successfully.

• Also included from the unpublished 0.5.1 draft: when the chosen model does not fit in memory, Ancilo tries a smaller context and then another local model that fits, explaining which one answered. It returns to the chosen model when memory allows.

Local model answers can still contain factual, language or planning mistakes. The chat changes do not validate every answer, and this release does not change your selected model.

Ancilo 0.5.0

• Sources you can open: an answer from your documents puts a small numbered mark after what it took from them – a click shows the very passage the answer had, with file, page or sheet and the text around it. Ancilo numbers the passages and checks the marks: a source the AI made up is taken out, and the answer says so. If the file changed since, the source says that too and still shows what the answer had.

• Look at a task's results before keeping them: tables and Word documents a task wrote are checked – does the file open, is what you asked for there (the columns you named, a total), do the totals and amounts add up – and Look at it shows the file with its rows and columns, a wrong cell marked. A problem is only what is sure; what Ancilo can only suppose (an empty cell, a total in the text that differs) is a note and never blocks a good file. Keep and Save take exactly the version you saw; if the task changed it since, Ancilo checks again first. The checks say what they cannot do: whether the content is right.

• Long tasks keep what matters: when a task's earlier tool output has to get shorter, a test run keeps its failures and its verdict (before: only its start – the failing test was the first thing to go), a list its first rows, a file its start and any line that stands out; each says whether the step failed. The whole output stays with the conversation, and the AI can read it again (read_result). With a small model (Qwen3.5 4B) on long multi-step tasks, 11 of 12 runs came out right instead of 7 of 12; on outputs Ancilo had never seen, the shortened text kept what mattered in 10 of 12 cases (before: 0).

• Fixed: the coding agent's commands failed on its own project path when its copy lies in a folder with a space (as Ancilo's data does, under "Application Support"): cd and file paths broke until the model found another way. Ancilo now writes the copy's path into a command the way its quoting needs.

• A computer too small for an AI model is told so, with the numbers – how much the smallest model needs, how much this computer can spare – and what would help, instead of "close other programs and try again". The setup goes on without a local model, and a chat says plainly why there is no AI to answer instead of offering a download that would not help. When only small models fit, Ancilo says what to expect from them.

Ancilo 0.4.5

• Updates you can see: when a new version is there, the header shows Update to 0.4.5 next to Manual – a click shows what is new and installs it with a restart (it warns if a task is still working). While Ancilo does not look by itself, the header has Check for updates (a little highlighted after a month without a look). The switch for the daily look moved to a box of its own, Updates – in System and in the setup – with the version and the last look. Ancilo looks at its start too when the daily look is on. Installing still always waits for your click.
• The setup checklist no longer has Ask something now – the Chat tab is right there.

Ancilo 0.4.4

• What left this Mac (System): everything Ancilo sent over the internet – web searches and the pages they read, model searches and downloads, the model list, messages to cloud models, update checks – in plain words: what, to whom, when, started by you or by Ancilo itself. A click shows each request in detail: the full address, what was sent, the answer. Kept 30 days on this Mac only, can be emptied; keys are never in it, and models (Claude Code, Codex, the assistant) cannot read it. Every request to the internet goes through one place that writes this log – a test makes sure nothing goes past it.
• Fixed: photos and scans had no text in the app. Text recognition worked from the command line, but inside Ancilo.app its sandbox kept macOS from reading the app itself – every picture came back empty. Found on a MacBook Air; now a receipt photo (JPEG, HEIC) and a scanned letter are read in half a second.
• A picture without readable text is said, not guessed: the AI tells you it cannot read it and asks for a sharper photo (before, a small model claimed to have no access, or guessed).
• Web search keeps your words: a small model sometimes misspelled a name in its search ("Gaustopfes" for "Gaustatoppen") or added a year you never named – now your spelling is used, invented years are dropped, and if nothing is found Ancilo searches once more with your own question.

Ancilo 0.4.3

• Two people, one Mac: when someone else on the Mac runs Ancilo too, Ancilo no longer fails to start – it takes the next free port (the other one's stays untouched), and never mistakes the other person's Ancilo for its own.
• No silent quitting: if Ancilo's background service cannot start, the app says why (in the Mac's language) instead of disappearing.
• Tables that open right in Excel: a CSV a task writes is checked before it is saved – amounts like 312,50 between comma-separated columns used to split into two columns. The model is told how to write it right (quotes, or ';' between the columns). Found on a MacBook Air with Qwen3.5 4B; now 3 of 3 runs give a correct table.
• Run from the download? Opened straight from the disk image, Ancilo asks to be dragged into Applications first – otherwise its background service would point at a place that is gone after ejecting.

Ancilo 0.4.2

• Security fixes from an independent review:
  • The coding agent's file tools no longer follow links out of the project. A broken link (to a file that does not exist yet) let write_file create a file outside it; grep and the file list read and listed files outside through links. Links inside the project work as before.
  • Ancilo's access token never reaches a model: Claude Code, Codex (MCP) and the assistant see it hidden – a cloud assistant asking for the model API settings got it. The app and the CLI still show it.
• License notices complete: the list now covers the native app too (Tauri and its components were missing), and Ancilo.app carries the notices itself (Contents/Resources/licenses).
• Dependencies are checked for known vulnerabilities on every change and weekly (just audit); the test runner is updated (vitest 4).
• Remove Ancilo (System › Remove Ancilo…, or ancilo uninstall): one click removes the app, its background service, the connections to Claude Code and Codex and – unless you keep them for a later install – the models Ancilo downloaded, conversations, settings and keys. Your own files, projects and the models of LM Studio or Ollama stay. Removing Ancilo Dev leaves the installed app's connection to Claude Code and Codex alone (both connect under the same names).

Ancilo 0.4.1

• Settings… (⌘,) in the Ancilo menu opens System; Help › Ancilo Manual and Manual in the header open the new user manual on ancilo.app (German or English).

• Ancilo's messages in German: errors and failed answers – memory, loading and downloading models, web search, documents, tasks, connecting Claude Code and Codex – appear in the app's language, also in conversations from before. Ancilo keeps them in English for itself (the history a model reads stays as it was).

Ancilo 0.4.0

• Fixed: "the model ended without an answer" on small reasoning models (Qwen3.5 2B on a MacBook Air: 8,192 tokens of thinking, 136 s, no answer). Chats now ask for the answer at once – seconds instead of half a minute – and any model that only thought is asked once more without thinking (chats, tasks and coding alike).
• You see what a chat is doing: loading the model, searching the web for “…”, writing the answer – with a clock after a few seconds, so a long wait never looks stuck. Tasks and coding show model loading and the clock too.
• Fixed: Ancilo Dev asked for the installed app's keys (a keychain dialog "ancilo wants to use your confidential information"). Each data directory now keeps its keys under a keychain entry of its own; the installed app keeps 'ancilo', so saved keys stay where they are.
• A tidy menu bar: only Ancilo, Edit and Window. The system's Services submenu (the services of every other app on the Mac), the empty File menu and Help are gone.

Ancilo 0.3.1

• Fixed: a loaded model was lost when its conversation grew – "needs about 41.6 GB, but only about 33 GB are free" on a Mac with 128 GB. Ancilo counted only "free" memory, not the file cache macOS hands back at once (where a stopped model's memory lands, too). It now counts what macOS gives a program, and a model whose larger context does not start always comes back as it ran (unless memory is critically short).

Ancilo 0.3.0

• A proper install window: the DMG opens a styled window – Ancilo on the left, Applications on the right, an arrow and one line in English and German between them; no toolbar, nothing hidden in sight (also with hidden files shown). Readable in light and dark mode.
• Pictures and scans have text now: photos of receipts (JPEG, PNG, HEIC, TIFF, WebP) and scanned PDFs are read by text recognition – the one built into macOS, on this computer, in a sandbox without network, at low priority. In chats, chat projects and tasks; recognized text is marked (it may have mistakes).
• Fixed: connecting Claude Code or Codex from the app failed with "cannot run claude: No such file or directory" – an app starts with a bare search path. Ancilo now finds them where your terminal does (Homebrew, npm, ~/.local/bin, nvm …), with the node npm tools need; if one is not installed, it says so and where to get it.