Changelog

Version 1.1.17 (2026-11-02)

New Features

Each MCP server has its own request timeout, set in its editor
AppleScript can control other apps, each app approved like a program

Improvements

Approved apps and SSH hosts have their own lists in Preferences › Security and in each agent’s settings
Plugin actions such as database writes and cleanups ask once, honor Allow All, and are refused in agent runs instead of waiting on a dialog

Bug Fixes

Text with a NUL character no longer loses everything after it when saved
A failed AppleScript is now reported as an error instead of a success

Version 1.1.16 (2026-10-25)

Bug Fixes

An iPhone reaches your Mac on the local network at once after reopening the app, instead of waiting 15 seconds

Version 1.1.15 (2026-10-25)

New Features

Agents can be added from the new gallery at cogitae.ai/agents and shared as documented agent files.

Improvements

The conversation history list loads faster.
Branches started, switched or renamed on the iPhone show up on the Mac, and iPhone messages stay on the branch the iPhone shows.

Bug Fixes

Clear completed in the agent panel no longer erases your agents’ run history.
The agent panel’s Clear completed button no longer looks active when there is nothing to clear.
The debugger tool can now explain why an agent run failed: runs, failed steps and tool errors are logged with their date and conversation.
The SQLite tool now queries the database you name instead of the last one opened, and no longer cuts long values short.
Editing an agent no longer drops AI settings the editor has no field for.
Agents given an instruction prompt now save and use its text.
Agents no longer pick a local model over your cloud models just because it is free.

Version 1.1.14 (2026-10-22)

New Features

The debugger tool lists what became of each request from the iPhone app (get_remote_log).

Bug Fixes

Switching branches is fast again instead of showing a spinning cursor.
Long conversations open faster.
Clicking a message on another branch in the conversation graph switches to that branch again.
A conversation whose first message has several lines shows as one line in the history list.
The Mac reconnects to the relay when its connection silently dies, so the iPhone no longer finds it offline.

Version 1.1.13 (2026-10-18)

New Features

Pair an iPhone over the relay with Pair iPhone over Relay in Preferences > Remote Access, by comparing a code as on your network
Preferences > Remote Access lists your paired iPhones with when each was last seen, and a Relay checkbox to keep one to your network

Bug Fixes

Only paired iPhones can now use your Mac through the relay: every message proves which iPhone sent it, and notifications go only to paired iPhones
Stop now ends the request to the AI provider straight away, so a stopped reply no longer keeps generating and costing tokens
Replies started by an MCP client (submit_query) can now be stopped from the chat
The MCP server now removes stored secrets from the conversations, prompts and other replies it sends to MCP clients
Exec approvals of programs outside Cogitae’s folders (Homebrew, scripts in your home folder) are now checked like any other program
Cogitae now asks before running a program it can’t check, and agents refuse it, instead of running it unchecked
Event agents with a tool list can now continue in a fresh turn, so Max continuations works for them
Sandboxed event agents with a tool list can now use the access tool to ask for a folder
An every-minute crontab schedule now says it isn’t allowed for agents and suggests “every 1m”, instead of reporting invalid syntax
Retry waits now follow a server’s Retry-After when it is given as a date
Agents now see the server’s suggested wait on every rate-limit error, so they wait or switch models more reliably
Large tool results reach the AI faster: secret scrubbing no longer stalls on very long lines or big outputs
Stop now also works when pressed before the reply starts streaming
Branching above a compacted point no longer sends the old branch’s summary to the AI
Conversations with an empty message, or a reply that only used tools, no longer fail with Anthropic
Attaching a .zip or other file a model can’t read is now refused with a message, instead of failing every later request
A request that fails on an unreadable attachment now names the attachment to remove
Images, audio and files attached to the system message are now reported as not sent, instead of being dropped silently
OpenAI models using the Responses API now receive text files attached to the system message
Importing a damaged or hand-edited conversation file no longer crashes Cogitae; files with broken message links are refused
Imported conversations keep messages marked “Hide from LLM”, and JSON imports keep their creation date
JSON conversation files from newer versions of Cogitae are now refused, as internal exports already were
Text after a NUL character in an imported conversation is no longer lost
Local MCP servers can return large results (images, big files) up to 32 MB; a larger reply fails with a clear error instead of hanging
Stopping a request to a local MCP server now ends it at once
MCP tools whose parameters use $ref, untyped values or bare enums are no longer dropped, and integer limits are kept
Reading resources from spec-conforming MCP servers, and prompts whose arguments omit “required”, now work
MCP tool results with newer content types (resource links, audio) no longer fail the whole call
Remote MCP servers that stream messages before their reply now work
Plugins with similar names no longer share stored data
Plugins asking for 0 or negative row limits no longer crash Cogitae or get every row, and conversation summaries show the real message count
Web pages converted to Markdown keep code indentation, nested lists, multi-line headings and escaped entities, and drop upper-case script text
Opening a long conversation from iPhone no longer drops a message beside the “earlier messages truncated” note
The Preferences sidebar keeps its width when you switch panes
Remote Access no longer shows your iPhone as offline because another paired iPhone is
Choosing branches from iPhone can no longer change branches in other conversations
iPhone requests for a missing or invalid conversation now get an error instead of no reply
Stopping an unknown agent from iPhone now reports “Agent not found”
Remote access now starts as soon as a license is activated, instead of after Cogitae restarts
The iPhone connection no longer drops while Cogitae shows a dialog, such as an approval request
A Mac clock up to a minute behind no longer rejects a just-issued license

Enhancements

Preferences > Security updates while open when a run approves or checks a program

Version 1.1.12 (2026-10-17)

New Features

Workspace agents now notice branch switches and new commits, even on a clean tree
Network agents now report whether an interface was added or removed, and notice joining another network

Bug Fixes

Any tool that would receive a secret (exec, ssh, MCP tools, plugins) now asks first unless the secret allows that tool; agents are refused
Secrets stored without their metadata are now checked before they are sent, like every other secret
Stored secrets no longer leak in base64 form when a tool reads a file with line numbers
Sub-agent results, event agent prompts and agent results pushed to the iOS app are now scrubbed of stored secrets
Exec approvals are tied to the approved program: if its file or its signer changes, Cogitae asks again
Approving a link now covers the file it points to, and the approval dialog shows both paths
Shells and interpreters (bash, python3, node, osascript…) are approved one exact command at a time unless you choose “Allow any command”
Fixed a crash when a local program sent the MCP server a malformed request
The MCP server now limits request size, checks its token before reading a request’s body, and refuses malformed Host headers
The Ask First approval dialog no longer freezes Cogitae while it waits
Approving or dismissing an Ask First agent after relaunching Cogitae no longer creates a second activation
Crontab timer agents now fire on monthly, yearly and leap-day schedules, and day-of-month and day-of-week combine as in cron
Crontab timer agents fire once when clocks change for daylight saving time
Calendar agents no longer fire twice when a meeting ends, and each occurrence of a repeating event triggers separately
Workspace agents now report DerivedData and .build changes, and no longer mistake files like rebuild-plan.md for build artifacts
Clipboard agents ignore copies made by Cogitae itself
Agent cooldown 0 now means no cooldown, and a timer agent whose cooldown equals its interval runs on every tick
A tool call that times out now ends on time, even if the tool never answers
A model calling a tool that doesn’t exist stops the reply with a clear error
Tool calls with malformed arguments are refused before they run
exec no longer passes the program’s own path as its first argument
Negative limits and offsets in tool calls are now rejected instead of returning every row or crashing
Tool number arguments sent as text or fractions are now read instead of ignored
Maintenance tools reject negative counts and sizes, and no longer crash on huge sizes

Enhancements

Tools that need a license now show “needs a license”, and the AI is told they’re unavailable
When two MCP servers offer the same tool name, the one listed first in Preferences is used, with a warning
Preferences > Secrets shows which tools each secret may be passed to
Preferences > Security shows what each approved program is bound to

Version 1.1.11 (2026-10-17)

Bug Fixes

Preferences editors now follow a switch between light and dark mode without a relaunch
Gemini no longer loses a second call to the same tool in one reply
Errors partway through a reply from Mistral, DeepSeek, llama.cpp, OpenRouter and custom servers are now shown, and retried when temporary
Importing a full conversation export now keeps its AI parameter and web search settings, including older exports
Importing a conversation now keeps which provider wrote each DeepSeek, MLX and OpenRouter message
Deleting a message with several replies no longer scrambles the order of its branches
Identical code blocks, tables, equations or images in one message now all show, not only the last copy
Bookmarks to headings, quotes, code snippets, images and checklists find their block again after the message changes
Opening a bookmark in a message with emoji no longer risks a crash
Agents falling back to another model no longer pick one without tool support when they need tools
A sub-agent whose parent model’s account was deleted now falls back to the default model
An MCP tool named like a Cogitae tool can no longer make that tool skip its confirmation; such MCP tools are no longer offered to the AI
Secrets are now redacted from tool output when JSON-escaped, percent-encoded or base64 wrapped into lines
Tool output no longer leaks a value whose name matches a stored secret with a different value
“Add to Secrets” no longer overwrites an existing secret with the same name
Passwords in connection URLs and in db-password= style lines are now masked in tool output
The exec tool now refuses more environment variables that make a program run other code; pager and editor variables may still be set to turn paging or editing off
The exec tool now masks sensitive values in command output as [REDACTED:NAME] instead of showing their first characters
Search no longer reads or lists files through links that point outside your workspace folders
Credential folders such as ~/.ssh stay protected however their name is capitalized
File tools no longer accept a broken link whose target is outside your workspace folders
Paths starting with ~ now always mean your home folder when checking folder access
Unattended batch file reads outside the workspace folders are now refused like single reads
The confirmation dialog now shows where a move, copy or conversion writes
The confirmation dialog hides more credential arguments, including camelCase and nested ones
Argument values in the confirmation dialog can no longer add lines of their own
Timer agents with different schedules now each run on their own schedule
File system and workspace agents watching different folders now each get their own folders’ events
Agent folder filters now include subfolders, and ** in a path pattern matches any depth
Triggering an agent with the agent tool no longer makes it run twice per event
An agent with an invalid schedule or folder no longer fires on another agent’s timer or folder
Notify Only agents started by the AI’s agent tool now send their notification
Agents waiting for a free slot keep their notification setting
An agent whose API key is rejected now falls back to another account instead of more models on the same key
Approved Ask First runs no longer leave a “pending” entry in the agent’s history, and run with the event that triggered them
Agent retries and fallbacks now get the same tools and full reply as the first attempt
Parallel tool calls and their results now stay in the order the model made them
Stopping a reply between tool steps no longer sends one more request to the model
At the 200-step tool limit, the last round’s tools are no longer run when their results can’t reach the model
A failed reply with many tool steps now shows one error instead of many
The MCP server now follows JSON-RPC 2.0 for batches, malformed requests and parse errors
Fixed a crash when an AI sent a very large number or a negative line limit to a tool
A number too large for a tool to use is now reported back to the AI instead of being quietly replaced by the default
Fixed Gemini rejecting tools with nested objects
Tool schemas now send default and enum values as numbers and booleans where declared
Plugin and MCP tool settings now show boolean defaults correctly
Plugin tools with untyped array parameters are now accepted by OpenAI
Cost estimates now use long-context pricing tiers
Built-in instruction prompts can no longer be edited or renamed by the AI

Enhancements

Agent active hours with the same start and end hour, such as 0–0, now mean all day

Version 1.1.10 (2026-10-05)

Bug Fixes

MCP clients are now told how to read the active conversation’s messages
Agents choosing a model by task type now pick the cheapest model that suits the task, not the highest scoring one

Version 1.1.9 (2026-10-01)

New Features

The agent panel has a button that drops down a list of agents to run now

Bug Fixes

An agent that gives up on its task shows as failed, not succeeded
A sub-agent on another model no longer inherits its parent’s Max Tokens
Agents keep their hippodamus plans instead of reporting “No active plan”
Agents running unattended no longer have their plans reviewed by Socrates
An agent’s token budget now counts the tokens of the sub-agents it starts
An agent retried on another model keeps its token budget and Notify Only limit
Models told to use Caesar, or to run a task on a named model, find the tool more reliably

Enhancements

The debugger can summarize a session’s tool calls as counts per tool
The cost tool reports a headless agent’s cost, sub-agents included
The agent editor explains each limit field, such as Max Tokens versus Token Budget

Version 1.1.8 (2026-10-01)

Bug Fixes

A new MCP server’s tools are ready to use in the open conversation and in new ones.
Copy Setup Command adds Cogitae to Claude Code in every folder, not just the current one.
Timer agents created by the AI or an MCP client get a working schedule.
An instruction prompt added by the AI or an MCP client appears in open conversations’ prompt menus.
The approval to run a program names who asks: the AI, an MCP client, or an app’s cogitae:// link.
Approving /usr/bin/git (or another developer tool) for an agent now lets it run that tool.
The AI no longer shows the folder picker for a folder it can already use, such as a Global Workspace.

Version 1.1.7 (2026-09-30)

New Features

Choose which tool calls ask first in Preferences > Tool Settings > Ask before
Anything that changes Cogitae itself (agents, MCP servers, plugins, tools, settings, secrets, grants) now always asks first
The approval dialog names who is asking: the AI, an MCP client, or the app that opened a cogitae:// link
The embedded MCP server now requires a token; copy the setup command from Preferences > MCP Servers
Pair your iPhone once to use it on your local network; every message on the network is now encrypted
Remote images in replies ask before loading from a new host (Preferences > Markdown > Remote images)
A web request sends a secret only to hosts you allowed for that secret
SSH asks before connecting to a new host
Notify Only agents now only read and report; they can’t change anything
Agents stop when they reach their token budget
Plugins not signed by BitArts ask before they load
Folder access can be read-only: check Read only in the folder picker
Click the tokens and cost in the status bar to see the usage report
Run now: a play button on each agent in Preferences > Agents runs it once right away
Agent results reach your iPhone: an agent set to Act and Notify or Notify Only sends its result to the Cogitae iOS app over the relay, even when the phone is away or the app is closed

Bug Fixes

Links opened by other apps or web pages can no longer run Cogitae tools without asking
Web pages can no longer reach the embedded MCP server
Deleted files now go to the Trash
Finder’s Run Agent on This opens the agent editor with the file, as intended
Sub-agents’ approval dialogs now appear in the chat that started them, instead of their operations being allowed or refused unseen
A tool that isn’t enabled in a conversation can no longer be run there
Global workspaces now work with the file tools after a relaunch
Removing folder access from the AI now reports what it actually removed
Deleting a conversation from anywhere now removes its folder access
Moved or renamed workspace folders are followed at launch
Secrets are hidden in tool output even if not used since launch, including their encoded forms
The approval dialog for commands no longer shows secret values
Commands that time out or are cancelled no longer leave child processes running
“Allow All Similar” now covers only the folder it was given in
Cogitae no longer quits when a daily timer agent is saved, or Cogitae starts, less than four minutes before the agent’s time
A HUD window moved away from the screen edges can be dragged back and docks again
The Set Dock Position shortcut and Focus filter now move the window right away
Charts of tall tables put the rows along the axis instead of drawing one series per row
Charts label their categories along the bottom, and bar charts with several series draw the bars side by side
Branches created over MCP appear in the conversation map right away
A branch created over MCP no longer sees the messages of the branch it sits beside
Branches created over MCP at the same time each keep their own model, and the conversation’s model is left alone
An instruction prompt set on a new conversation before its first message now sticks
A LaTeX equation that cannot be rendered shows its source instead of a blank gap
A table that cannot be charted no longer switches your default table style back to Table
Notes added and messages moved over MCP show in the open conversation right away
The status feed shows one copyright notice at startup, with the current year
An edited message’s branch label matches the branch picker
A conversation created over MCP shows in the conversation tool’s list before its first message
The status bar badge counts only warnings, errors and other notices worth a look, not every query’s progress notices
The message input no longer shows a scrollbar after attaching a file, and returns to its empty layout after sending
The AI can read the text of PDFs in a workspace folder directly, instead of writing scripts to extract it
“Allow All Similar” in the approval dialog now also covers the operations already waiting, so sorting a folder asks once
A question the AI asks you shows in full, scrolling when it’s long, instead of being cut off
Adding a workspace over MCP or by the AI asks you to grant access to the folder, so its tools can read it
A branch answered by another model is told which model it is, not the conversation’s model
Access to a folder no longer counts for a folder whose name only starts the same way, and approving a program no longer approves others whose names start with it
Sorting or copying many files takes one step instead of one per file, and uses far fewer tokens
A tool no longer times out while its approval dialog waits for you
A DeepSeek API key pasted in onboarding becomes a DeepSeek account, and a key the provider rejects is named instead of skipped silently
Pasting an API key into a new account and clicking Save no longer says “Please enter an API key”
Accounts can be deleted with a button in the account form or by right-clicking them, not only by swiping
Opening an account’s settings no longer widens the Preferences window, and scrolls the whole form into view
The sidebar’s conversation graph fills the sidebar’s width and opens on the newest messages, and clicking a node goes to that message
Messages hidden from the AI stay hidden from its tools, from MCP clients and from the conversation’s title, not only from the chat history it’s sent
The Calendar permission prompt explains that the AI reads your calendar when you ask about your schedule, not only that agents use it
Typing a new template’s name replaces the default name instead of going in front of it
Text fields follow your Keyboard settings for spelling correction, text replacement and smart quotes instead of always correcting
Copy & Paste Back pastes the Launcher’s answer into the app you were in
Opening the Launcher no longer brings Cogitae’s main window forward
The Launcher hotkey opens the Launcher on the first press even right after it was closed
The help gives the right default for the inline text hotkey (Control+Command+Space) and says to turn it on first
Conversation titles no longer show markdown such as asterisks
The bookmarks sidebar shows the date under each bookmark’s name, so the name gets the whole row
Preferences opens taller the first time, with room for the agent editor, and keeps the size you give it
A timer agent no longer gives the same answer every run
The AI follows remembered preferences, such as a diet or metric units, more reliably
The usage report shows its cost totals on separate lines
Questions about model prices use about half the tokens, and “cheapest” answers skip weak models
When memory injection is off, the AI says so after saving a memory instead of promising new conversations will know it
The AI saves personal details you ask it to remember, such as a food allergy, instead of refusing them as sensitive
After Copy & Paste Back, the keys you type next no longer act as Command shortcuts, and the Launcher hotkey keeps working
Copy & Paste Back pastes the answer as the Launcher shows it, without markdown symbols or extra blank lines
A conversation opened from the Launcher shows the tokens and cost its answer used
A single line break in an answer, as in “Warmly,” above a name, is kept instead of running the lines together
The usage report explains when cache writes cost more than they saved, and its 30-day totals count all conversations
The MCP server retries when its port is briefly in use at launch, and reports the port it listens on
Searching memories finds them by a single word or a short phrase, such as a project name
The calendar tool accepts a date and time without a time zone as your local time
Research reports from Newton and other agents open at a readable size instead of a tiny window
Links in footnotes, such as a source’s DOI, are clickable instead of showing as markdown
Footnote markers side by side, such as [^1][^19], read “1,19” instead of running together as “119”
A code block titled file=name shows the file name without the asterisks, with its language
The conversation graph shrinks to show several branches side by side, and Shift with a mouse wheel zooms it
Switching to a branch switches the model picker to the model that branch used
The AI can run git and python3 on a Mac with only the Command Line Tools, and finds the real binaries
A command the AI runs that fails to start, such as one from Homebrew, reports why instead of hanging Cogitae
An agent’s log says why a model declined its request, and an agent finished by a fallback model logs as completed
A plugin’s confirmation dialog, such as Empty Trash, no longer times out the tool while it waits, and one approved after a cancel does nothing
The Maintenance Toolkit’s cleanup lists what it will delete, empties cache and log folders instead of failing to trash them, and leaves /tmp alone
The status bar says “Running 1 tool” instead of “Running 1 tools”
The AI can run programs you approve from Homebrew or that you built yourself, through the same launcher local MCP servers use

Enhancements

The License pane and plugin badges say Paid instead of Premium, to match the Paid license you buy

Version 1.1.6 (2026-09-27)

Bug Fixes

Setup now chooses summarization and embedding models instead of leaving them unset
The iOS model setting can follow the Default Model
Tables with text columns before the numbers can be shown as charts
Chart axes show large numbers as 2M instead of 2.0E6
Branch menus no longer go blank when two branches start with the same words
Branch menus show the branch name, the same as the label on the branch
The conversation graph shows a branch’s name along with its message
Local MCP servers started with npx, uvx or Homebrew now run, after a one-time launcher install
Adding a second MCP server that uses the same command (such as npx) no longer fails
MCP tools whose arguments accept more than one type (such as DeepWiki’s ask_wiki_question) are no longer dropped
Agents can use MCP tools; before, they were told the tools were unavailable
A finished agent no longer disconnects MCP servers that are still in use
Agent cards show the tool the agent last ran
Replies to queries sent over MCP show formatted, without switching conversations
Conversations created over MCP no longer start with a “New conversation” note
MCP server lists say “1 tool” instead of “1 tools”
Conversation graph previews show plain text instead of markdown symbols

Enhancements

Local server models are shown by file name instead of full path

Version 1.1.5 (2026-09-27)

Breaking changes: updating resets the old provider setup, so enter your API keys again in Preferences > AI. Shortcuts and Focus filters now pick a model instead of a provider, so choose one again in each existing shortcut.

New Features

Accounts replace provider entries: one per API key, and every model it serves is usable in any conversation
New conversations start on a default model chosen for you, changeable in Preferences > AI
The model menu groups models by account, with More Models and a search for every model an account serves
Preferences > Models lists every model with its scores and price, which ones agents may use, and each model’s settings
llama.cpp models are recognized by their weights and take catalog scores automatically
Agents can be limited to chosen accounts and a maximum price, and their sub-agents keep the same limits
The new-model notification has an Enable for Agents button
Templates remember the model they were saved with, not only the provider

Bug Fixes

Fixed llama.cpp conversations failing on their second message
Fixed a sub-agent failing when its model request was left blank
Fixed the template tool’s apply action doing nothing
Fixed an error in the status bar after finishing onboarding
Fixed Newton research being declined by Claude models partway through
Fixed a failed research agent being reported as finished
Fixed the first conversation after onboarding starting on an arbitrary model instead of the default

Enhancements

Model lists come from the model catalog and show each model once
Cost estimates use a model’s long-context prices when a prompt reaches them
Onboarding masks the API keys you paste
The embedding model can be chosen from each provider’s embedding models
Onboarding recognizes Anthropic, OpenAI, xAI, Gemini, OpenRouter, Mistral and DeepSeek keys, and ignores labels around pasted keys
Agents choose among the models enabled for them, with scored local models competing at no cost and local models as the last resort
A rate-limited agent switches to another account instead of leaving the provider, and tries up to three other models
Agents skip a llama.cpp server that is down or cannot run tools
The Extended agent model pool setting is gone: enable models for agents in Preferences > Models
The cost tool prices local models and tells apart models with the same name on different accounts
Newton and the other research agents are no longer started for simple questions
An agent whose request a model declines retries on another provider

Version 1.1.4 (2026-09-26)

New Features

Agent cards show the total cost alongside tokens once an agent completes.

Bug Fixes

Fixed “prompt is too long” errors after the AI fetched a large web page.
Fixed the status bar’s token and cost totals shifting left when an error is shown.
Fixed a crash when saving a reply after the AI took a screenshot.
Fixed Claude agents failing partway through when a tool search ran alongside other tools.
Fixed the input field’s placeholder naming a different model than the one the conversation uses.
Fixed agents stopping after asking to continue when their reply had no text.
Fixed agents losing their conversation when continuing to the next turn.
Fixed sub-agent and ask-user waits being cut off after 60 seconds.
Fixed the Caesar agent card staying “running” after all of its agents finished.

Version 1.1.3 (2026-09-24)

New Features

Live progress while a reply is prepared, including model reasoning summaries, on Mac and in the iPhone app.
The AI now remembers the tools it called in earlier replies, their arguments and results, even across provider switches.
Expanded web search: Claude and Gemini can read web pages, Grok searches X, OpenRouter models can search, and OpenAI, Gemini and Grok replies list their sources.
PDFs are sent as files to models that read them (Claude, Gemini, current GPT), so layout, tables and scans come through.
Declined Claude requests are retried on a fallback Claude model.
Warnings for retired and soon-to-retire models, with a suggested replacement and an offer to switch.
New processing options: Verbosity for GPT-5 and later, Priority processing per conversation, and half-price Flex processing for agents on OpenAI and Gemini.
Reasoning effort settings for OpenRouter, llama.cpp, Mistral and DeepSeek, listing only the levels each model offers.
Replacing a Mac no longer requires the old one: deactivate any of your activated Macs from the new one.
Preferences > Plugins now shows plugin updates and offers a restart when one is needed.
MCP clients can attach files and switch or delete conversations and messages.

Bug Fixes

Fixed requests failing on newer Claude, GPT, Gemini, Grok, Mistral and DeepSeek models when reasoning, sampling, token or stop settings were set. Cogitae now sends only what each model accepts.
Fixed provider-specific failures: Claude tool calls with thinking on, Claude Fable 5.1 declining every request, GPT overwriting optional tool settings, OpenAI rejecting some tools, Gemini blank replies and web search with tools, DeepSeek empty replies, and Mistral ignoring audio and JSON schemas.
Fixed local and custom-server models not being offered tools, custom scoring needing a restart, and a blank context window being treated as 8,192 tokens.
Fixed token counts and costs across providers: overstated cache costs, under-counted tool loops, and failed, retried or background requests missing from totals.
Fixed OpenAI and Grok replies silently stopping when the provider was overloaded; they now show the error and retry.
Fixed automatic compaction starting early after long tool runs or web searches.
Fixed agents failing on retired models, gaining tools outside their setup, losing folder access after a retry, and losing model settings on restart.
Fixed agent and background replies overwriting the status bar for the conversation on screen.
Fixed Cogitae removing your license key when the server did not recognize it, and a released Mac staying licensed.
Fixed the iPhone app waiting forever on failed or unreachable Macs, listing your Mac twice, and losing the local network connection.
Fixed the embeddings plugin ignoring your Embedding AI provider, and re-embedding leaving entries behind on rate limits.
Fixed Claude web searches stopping partway through and text around a search running together.
Fixed file and screenshot tools saving outside Cogitae’s data folder.
Fixed new conversations sometimes starting with the wrong provider, and the API endpoint setting being ignored.
Fixed Gemini context caches staying on Google’s servers, and billing, after quit.
Fixed unusable models being offered and retired models reappearing in menus.
Fixed a window resize when switching providers, a crash setting stop sequences, and misleading error advice.
Fixed the image generation plugin listing models more than once (image generation plugin 1.1.4).

Enhancements

Longer conversations cost less: Claude, GPT-5.6/GPT-6 and OpenRouter conversations stay in the prompt cache between messages, and the tool optimizer no longer discards it.
Claude models on OpenRouter use Anthropic’s format, with visible thinking and Claude’s own web search.
Gemini 3 can see images returned by tools, audio attachments reach Gemini, and dropped attachments show a warning.
Token usage now shows exact Grok and OpenRouter costs, DeepSeek off-peak pricing, and embeddings, images, web searches and Gemini caches.
Clear messages when a Claude model declines a request or a reply fills the context window.
Updated defaults: current fallback model lists, Gemini Embedding 2 with automatic re-embedding, and GPT Image 2 for image generation.
The Account panel is now called License and shows Mac slot usage.
The ai_config tool can set every reasoning, sampling and output parameter.

Version 1.1.2 (2026-09-21)

New Features

Added automatic update checking — Cogitae now detects new releases and lets you know when one is available.
Added agent start and stop dates and times to agent cards.
Added status bar notifications when new models become available for your configured providers.

Bug Fixes

Fixed tool selection expand/collapse in the agent editor.
Fixed embedding mismatch in the Embeddings tool when the embeddings provider changes.
Fixed timer agents firing multiple times for a single scheduled event.
Fixed popup menu items becoming invisible when highlighted; text now uses the label color instead of the accent color that also fills the selection bar.

Enhancements

Added a key legend showing the shortcut for submitting a query.
License tokens now identify the product they were issued for, so a key issued for another product is rejected with a clear message instead of a generic failure.

Version 1.1.1 (2026-04-01)

Bug Fixes

Fixed agent card elapsed time continuing to increment after agent completion.
Fixed tooltips not showing in dashboard buttons.
Fixed editor freezing when deleting items from the executable list in agent preferences.
Fixed layout bugs in the agent preferences editor.

Enhancements

Optimized memory system usage with improved embedding service and maintenance routines.
Rearranged conversation history context menu items to a more logical ordering.
Adjusted which tools need a Paid license or an active trial.

Version 1.1.0 (2026-03-28)

Major New Features

Autonomous AI Agents — Event-driven agents that monitor your system and respond to file changes, clipboard activity, timers, calendar events, network changes, and more. Agents run headless AI sessions with configurable action policies (notify only, ask first, act silently, or act and notify). Includes a dedicated Agent Panel for monitoring and control.
iOS Companion App — Remote access to your conversations from iPhone and iPad via a WebSocket relay service. Includes onboarding flow, QR code license scanning, demo mode, and remote agent management.
MCP Server — Cogitae now hosts an embedded Model Context Protocol server, exposing all tools, conversations as resources, and prompts to other AI applications on your machine.
Memory System — Long-term memory with semantic search, automatic extraction from conversations, trust scoring, decay management, and a Memory Admin tool. Memories are injected into prompts when relevant.
Local AI with MLX — Run models locally on Apple Silicon using the MLX framework. Includes a model downloader from Hugging Face with size estimation, model swapping, and in-process inference with no external dependencies.
Spotlight-Style Launcher — A global hotkey activates a system-wide launcher for quick access to Cogitae from anywhere, with inline text triggers and voice input.
Plugin System — Seven bundled plugins: AppleScript, Embeddings, Image Generation, Maintenance (23 system tools), Plugin Builder, PostgreSQL, and SQLite.
App Intents — Siri and Shortcuts integration for controlling Cogitae from outside the app.

New AI Providers

DeepSeek
OpenRouter
xAI image editing support

New Tools

Sherlock — Investigation and research tool with programmatic fan-out for deep analysis.
Patton — Security assessment agent with recon-driven threat discovery.
Caesar — Sub-agent coordinator for spawning and managing independent AI sub-tasks.
Hippodamus — Planning tool with save/load support for persistent plans.
Davinci — Innovation and creative problem-solving tool.
Newton — Research and scientific analysis tool.
Aristotle — Deep exploration and analysis with domain-specific disciplines.
Socrates — Critical review and Socratic questioning tool.
Dashboard — Visual dashboard panels for presenting structured results.
Clipboard — Read and write system clipboard contents.
Screenshot — Capture screen contents for visual analysis.
Random — Random number, float, and pick generation.
Show File — Direct file content display as visible Note messages.
Read — Read-only file access for sub-agents that investigate but never change files.
Memory Admin — Manage, search, and maintain the memory system.
SSH — Remote filesystem access and command execution over SSH.
Batch Read — Concurrent multi-file reading in a single tool call.
Tool Optimizer — Dynamically adjusts available tools based on conversation context.

Enhancements

Added extended thinking support for Claude Opus.
Added web search support for Anthropic provider.
Added structured JSON output support.
Added image URL support in conversations.
Added Google-specific API features including context caching.
Added Table of Contents sidebar with collapsible sections and interactive graph view.
Added bookmark navigator with fuzzy search and custom icons.
Added prompt template system for saving and reusing conversation setups.
Added auto-continuation when context fills up during long tasks.
Added token and cost tracking per conversation with per-provider pricing.
Added hide-from-LLM toggle to exclude specific message pairs from AI context.
Added pinch-to-zoom in conversation graph view.
Added message type indicators showing the kind of each message.
Added inline expand/collapse animations in editors.
Added swipe-to-delete across all list views (conversations, bookmarks, agents, etc.).
Added drag-and-drop import/export for agents and MCP server configurations.
Added per-tool timeout settings.
Added per-agent reasoning effort control for sub-agents.
Added composable prompt fragment system with built-in presets.
Added conversation summaries that update dynamically.
Added copy-to-clipboard in conversation history context menu.
Added configurable network connectivity during system sleep.
Added Mermaid diagram dark mode support for printing and export.
Refined Preferences UI across all panels for consistency.
Improved branch navigation and tree-structured conversation management.

Security

Sandboxed application with security-scoped bookmark inheritance.
Added executable security verification for tool execution.
Added SSRF protection for web fetch and image rendering.
Added secret detection dialog when unrecognized secrets are found.
Added global workspaces with drag-and-drop in Security preferences.
Google API key moved from URL query parameter to request header.
Relay authentication token moved from URL query parameter to Authorization header.
Secrets are now redacted in query lifecycle logging before persistence.

Bug Fixes

Fixed inline LaTeX parser misinterpreting currency symbols (e.g. “$3.00 of your $10”) as math expressions.
Fixed race condition in continuation tool execution.
Fixed security-scoped bookmark access for workspace files.
Fixed conversation branching, reparenting, and deletion preserving child messages.
Fixed chat scroll sticking and bouncing issues.
Fixed multiple AI responses when continuing a conversation.
Fixed agent panels losing content after sub-agent completion.
Fixed token counting to correctly sum input tokens and prefer per-model pricing.
Fixed cache token double-counting in cost calculations.
Fixed duplicate dashboard panels stacking on repeated tool calls.
Fixed a crash in markdown rendering.
Fixed AI provider editor removal race condition.
Fixed continuation instruction messages not laying out properly.
Fixed display issues in the Conversation History window.
Fixed scheduled agents not firing during system sleep or App Nap.
Fixed various sandbox-related path issues for home directory access.
Fixed transient provider error retries with exponential backoff.
Fixed status bar token count to include both prompt and completion tokens.

New Plugins

Embeddings Plugin — Local semantic embeddings using Natural Language framework.
Image Generation Plugin — AI image generation with xAI editing support.
Maintenance Plugin — 23 system maintenance and cleanup tools.
Plugin Builder Plugin — Create new plugins from within Cogitae.
PostgreSQL Plugin — Direct PostgreSQL database access.
SQLite Plugin — SQLite database browsing and querying.
AppleScript Plugin — Execute AppleScript automation from AI tools.

Version 1.0.2 (2026-02-06)

Bug Fixes

Fixed tool settings not appearing dynamically when a tool is installed
Fixed UI shifts when changing App Mode: changes take effect on restart
Fixed images silently dropped when using Anthropic or Gemini providers
Fixed MCP paste JSON panel text invisible (text view sizing issue)
Fixed MCP server discovery failing silently after successful connection
Fixed MCP tool list not updating in chat after server discovery
Fixed MCP server preferences showing stale tool counts after save
Fixed MCP resources/prompts UI not appearing
Fixed tables truncating text

Enhancements

Handles long plugin descriptions more gracefully in Plugins Preferences
Remove “META” tagging for meta messages
Claude Opus 4.6 support (adaptive thinking, effort setting, 128K output, interleaved thinking)
Anthropic web search (domain filtering, citations, status bar updates)
Structured outputs (strict tool use, JSON schema output mode for Anthropic/OpenAI/Grok)

Non-functional Changes

Added note about summarization backend setting

New Plugins

Maintenance Toolkit — cleanup tools for deleted app artifacts, duplicates, etc.

Version 1.0.1 (2026-02-04)

Bug Fixes

Fixed growing input field padding
Fixed plugin parameters not persisting

Enhancements

Added About panel in preferences

Non-functional Changes

Updated copyright date

Version 1.0.0 (2026-02-01)

Initial release

Last updated November 2, 2026