What shipped, and when

A new shape for a map that had got too wide: the same hierarchy hung off a trunk, so it runs down the screen instead of across it. And a run of repairs to exporting a map as a file — large maps, long blocks, names in other alphabets, and a browser with nobody signed in.

What changed

• Spine, a new shape in the Arrange menu. Each block drops a line and its branches hang left and right of it, two to a rung, rung after rung downward — so a map with a lot of branches stays narrow enough to read instead of turning into a strip.
• Every block does it, all the way down, so a deep map does not fan out further at each level. Small groups still look ordinary: an only child sits straight under its parent, and an odd last one sits on the trunk between the others.
• Exporting a map as a file: a large map now reaches PDF. Past about thirty-five blocks the export used to stop and report an error instead of producing anything — which is the size at which a map is worth printing. It produces a file at any size now, and every page after the cover carries its number at the foot.
• A long block in a slide deck continues onto the next slide instead of printing over the boxes beneath it. Each continuation keeps the same number, heading and colour band and is marked (cont.), and a numbered list that carries over picks up where it left off.
• A map named in Russian keeps its name in the downloaded file, rather than every such map arriving as the same untitled one. One slow photograph can no longer hold an export open with nothing on screen to say so.
• A big export no longer freezes the page. The map stays usable while the file is built, the menu says what stage it is at and how long it has been going, and Cancel stops it. Closing the menu does not cancel anything — the file still arrives, and reopening the menu shows the same run still counting.
• The fonts a PDF export needs are served whether or not there is a session. A PDF carries its own typefaces, so that a map written in Russian comes out as words and not as empty boxes, and the browser building the file fetches them from here; a request arriving without a session used to be answered with the sign-in page instead, and the export then died on a typeface it could not read.
• Hiding a map's dashed links now reaches the file you download: the exported document leaves those cross-references out and says on its cover how many it did not list. What your assistants read is unchanged.

The assistant's reasoning can be left open while you go on using the map: a menu no longer shuts itself every time that panel writes another line.

What changed

• With the assistant's thinking expanded, every dropdown on the page — the chat's mode and branch pickers, a toolbar menu, a map row's ⋮ — used to close itself several times a second while the model thought, and took the keyboard out of the menu with it. They stay open now.
• A hover label stays up while the assistant is thinking, instead of disappearing the moment it writes another line.

An assistant you connect can now hand a whole map to a second assistant that builds on it, and work handed over that way can take minutes without the connection dropping.

What changed

• A connected assistant can now hand one of your maps, whole, to a second assistant that can add blocks to it — so a document put on a map as one block can be turned into a map. Until now an assistant handed part of a map could only rework blocks that were already there.
• A map you were only invited to cannot be handed over whole. Hand over one branch of it instead.
• Work handed to an assistant now keeps its connection open while it runs, with a progress update about every 20 seconds, so a job that takes minutes reaches the assistant that asked for it instead of failing on the way back.
• One handed-over job now stops after five minutes and says so. Everything it had already changed stays on the map.

The chat on a map now works in modes — Builder, Writer, Reader and Orchestrator — and each one can do only its own kind of work.

What changed

• Pick a mode from the chat's header. Builder is the chat as it was. Writer fills in the words of a map that already has its shape. Reader answers, reviews and changes nothing.
• When a request fits another mode better, the chat suggests it with a Switch button and a note to carry over. Nothing runs until you press Send.
• Reader's suggested edits, moves, deletions and links each have their own Apply, and an applied one goes into undo like a change made by hand.
• Orchestrator hands tasks to the branches of the map you are on and reports what each one did, while the canvas updates as the work lands.

The model list now matches what each provider offers today, and a model it does not carry can be typed by its ID.

What changed

• Every plan now runs on DeepSeek V4.1 Flash with no key of your own. DeepSeek retired V4 Flash, V4 Pro is no longer offered, and a map, bot or key set to either now runs on V4.1 Flash.
• The list now has Claude Fable 5.1, Claude Opus 5, GPT-6 Astra, Gemini 3.8 Flash and Gemini 3.5 Flash-Lite in it.
• With your own key for a provider, you can type any of its model IDs in Settings ▸ AI Chat. The chat's Model menu links there.
• In the chat menu's skill list, “Always in context” now sits under the skill it belongs to. It no longer looks like a skill of its own.

A map you publish now works in full for the people you send it to, whether or not they have an account.

What changed

• Someone opening a published map without signing in now sees its pictures. Until now every picture on a shared map was a broken image for exactly the people the link was for.
• A map shared by link can now be shown working inside another page: visitors drag blocks, draw connections, open them, switch to the graph and download the map, and nothing they do is saved — reload, and it is the map you published.
• The page a map is shown in can now steer it — hide the toolbar, switch to the graph, frame one block, light up a branch while the rest dims, start over — and the map answers every request and tells the page when a visitor first touches it.
• A map shown on another page can now show a visitor, once, how to connect two blocks: a cursor drags a link from one block to the next, then leaves the map to them. Their first touch stops it on the spot, and for anyone who asked their system for less motion the link simply appears.
• A map shown on another page can now play a recorded interview beside it: the questions print in the app's own question form, the answers type themselves, and the map builds itself — then it is the visitor's, from their first touch. No model is asked anything and nothing is saved.
• A map shown on another page can now replay a real conversation someone had with its agent: the messages arrive in the app's own chat panel, and the map beside it redraws to each state that conversation left it in — on the graph view when the agent was the graph analyst. The visitor's first touch hands them the map as it stands.
• A map shown on another page can now be seen as the people on it see it: the page switches between the whole map and one person's branch, the rest of the map leaves the screen, and switching back brings it home — plus what an agent seated on that branch will read, marked Soon.

When the assistant finishes a turn, or a screen fails, you get a short reference you can quote. The people who run this deploy can see what is broken, and a schedule now says whether it worked — not only that it ran.

What changed

• Every conversation with the assistant now leaves a record you can point at: a quiet eight-character reference, and behind it the tools that ran, in order, with how long each took and the name of anything that went wrong. A turn that finished and built the wrong thing used to leave nothing behind at all.
• When something fails — on your maps list, on a map that would not open, in the assistant, or in Settings — the banner now prints that kind of reference. A support request you file from Settings carries one automatically, together with the app’s version, and nothing else about you or your maps.
• A page that fails to render now reports itself instead of writing into a browser console nobody reads.
• A schedule now says which of the four it is, in words: never run yet, worked, failed with the reason, or paused after too many failures in a row. A run that worked used to look the same as one nobody had ever checked.
• For the people who run this deploy, /admin now answers “what is broken” and “what does the assistant actually do”: spend the ledger failed to record, models billed at an unknown rate, failures grouped so the same fault two hundred times reads as one row, whether each background job is still being called, which tools are used and which have never been. Type in a reference from someone’s screen and the run opens, tool by tool.
• When a threshold is crossed, the deploy’s owner can be messaged on Telegram with a rule name, a count, a rate and a window — no account, no map, no error text. Until a chat is named, the watching is built and silent.

Pictures and workspace attachments now live in the object store beside recordings. Postgres no longer holds any data columns, and every byte is reclaimed when a board or an account is deleted.

What changed

• The application now has one answer to “where do bytes live”: the object store. All three byte kinds — pictures, workspace attachments and recordings — share a unified port and the same account budget.
• Postgres holds no bytea columns any more. Drawings ride along for free, since a drawing is an image.
• Deleting a board or purging an account now cleanly reclaims its objects from the bucket. The logic is shared across all three stores and ensures nothing is left behind.
• Serving stays proxied through the application, so your permissions are still the one true authority and no signed URL ever leaves the deploy.
• Search learned a third question. “By word” finds your exact words inside every map, quotes the line they are written on and marks them in it — free, instant, and answering for a block you saved a second ago. Pressing a result opens that map framed on the block, and the board’s own search now reads the blocks themselves, not just their titles.

A new account no longer starts at an empty grid: up to three ready-made maps are already there, they are yours to edit or delete, and none of them counts against your map limit.

What changed

• Sign in for the first time and up to three ready-made maps are waiting, chosen for the Work field on your Corlecti account. They are yours — rearrange them, rewrite them, delete them — and each card says so with a small “Gift” marker.
• None of them counts against your map limit, ever, including after you have edited one beyond recognition. Deleting a gift frees nothing, because it never took anything.
• A picture inside a gift shows up without costing you a byte of your own storage.
• The Corlecti roadmap arrives under your shared maps, read-only. It is not a copy but the live map we keep, so it is current every time you open it — and if it is published later, it reaches your account at your next sign-in.
• “Start your first map” is back for anyone who has never made one: it now counts the maps you authored, so a brand-new account sees the gifts and the panel together, and the panel goes away the moment you make a map of your own.

Access, skills and assistants now follow a folder down to the maps inside it and say so wherever you meet them — and the workspace chat says who else is on a map and whether what you sent has been read, both derived from what was already stored.

What changed

• A map that sits inside a shared folder is now marked “Shared folder” rather than “Workspace”, and the mark names the folder the access comes from. The two are different facts: people really can change such a map, but it is not a shared workspace of its own — and adding your first editor there still spends a workspace slot, because a slot is spent where the invitation is written.
• That refusal now names what is already holding your slots, and says when one of them is a folder above the map you are on: the people on it can already open this map, and making this map a workspace of its own takes another slot. It used to be a bare “you have used every shared workspace on your plan” beside a map already marked as shared.
• Map settings → Skills shows a skill reaching this map from a folder above as switched on, locked, and captioned with the folder’s name, so “why is the assistant doing that?” has an answer on the map itself.
• The same row in the chat panel’s Skills list reads the same way, and no longer offers a tick that could not change anything.
• A skill set to “Selected maps and the maps inside them” is described that way everywhere it appears. It used to be captioned “Available on every map.” on the one screen that explains what the tick does.
• An assistant bound to a folder can now run a scenario on any map inside it. It used to be able to start one on a sub-map and then reach neither the read nor the step that continues it, so the interview could not be finished or closed.
• A search now says which of your maps it looked in even when it finds something — a search narrowed to sub-maps or to one assistant’s branch used to return a bare list, which an assistant could present as everything you have ever written.
• list_agents says how far each assistant reaches, not only which map it names, so a key covering a folder is no longer described as covering one map.
• Someone invited to a folder can now open a map inside it that is published to “only invited people”. That link used to check the people invited to that one map, so it refused the very people who open it every day from the map list.
• “Applies to” on a person’s row is now shown wherever that is what their access says — not only while the map still holds sub-maps. Moving the last map out of a folder used to leave the setting on “This map and the maps inside it” with nothing on screen saying so and no way to narrow it.
• Opening a link into a folder no longer opens by saying the folder is gone. The page said so before it had fetched your maps — it could not find the folder in a list it did not have yet — so every reload and every shared link into a folder started with a line about your own maps that was not true. It now waits until your maps have arrived.
• Deleting a map counts its sub-maps the same way from the map grid and from the drawer, reading the count that comes with the map rather than counting again from whatever part of your list that screen was holding.
• Moving a map now names everything that will stop reaching it — people, keys and skills. It used to count only people, so taking a map out of a folder that seats a key changed what an assistant can do with no warning, and skipped the confirmation altogether. When what you are moving is a folder, the sentence says so about the maps inside it too.
• A folder’s reach is worded the same way everywhere you meet it — on a person’s row, on a skill, on an API key and in the chat — and the API-key form’s picker is headed “Applies to”, like the one beside a person. The four wordings said one thing four ways, and the one that survives says “the maps inside it”: a scope reaches maps, and “everything” left you wondering what else.
• The panel’s header is one strip: «Everyone», then a chip for each person on the map. Press somebody to open the conversation with them; press again and the same conversation comes back rather than a second one.
• A person’s chip carries its own unread count, and somebody nobody has written to yet simply has no badge.
• Your own messages now say whether they landed: one tick when sent, two when everybody they went to has read them.
• In the map’s channel a «Read N» sits beside the tick — press it to see who, by name.
• The unread mark on the chat control is red instead of green, so it reads as something waiting rather than as one more control that is switched on. It follows the control into the ⋮ menu if you have tidied it off the toolbar.
• The separate «Message» picker in the panel’s header is gone. The strip does both jobs, so starting a conversation and returning to one are the same press.

Each Telegram bot you registered, and each key you create, can now run on its own model and its own reasoning level — and a bot can say whose key pays for it.

What changed

• A bot’s card has one more group, closed until you open it, holding the model it answers with and how hard it thinks. Both start at «use your account’s», which is exactly how every bot behaved before.
• You can also say which key a bot runs on: your own when you have one, so your provider bills you and the turn costs nothing against your plan allowance — or always this deploy’s shared key, which counts against your plan allowance instead. Both options say where the bill lands, next to themselves.
• Forced onto the shared key with a model this deploy does not cover, the card says which model your turns will actually run on. It never falls back to your own key.
• A new API key can be created with a model and a level too. They are used when an assistant hands work to another one, and by nothing else — every other request from a connected client is answered by that client, on its own model.
• The key list now says what each key was created with, because a key cannot be edited afterwards and that was the only moment you could ever see it.

The chat panel keeps its settings in one menu instead of a strip above the composer, and the dropdowns stopped being able to take the page down with them.

What changed

• The preset, the assistant’s layout, the palette, the scenario you could start and the skills in force are now in the chat’s ⋮, beside the model — one press away, and the composer gets its space back.
• Opening and closing a menu quickly no longer freezes the page: a panel re-opened while it was still closing could get stuck correcting its own position until the tab gave up.
• Opening a map no longer leaves an error behind in the browser console.

An assistant seated on one branch of a map can now be given a job by another one — so a map divided into branches can be worked on by the assistant that belongs to each, from one conversation.

What changed

• A connected client can hand a task to any assistant you have seated on a branch, and gets back what it did and what it said.
• Each one stays inside its own branch, whatever it is asked for.
• Two assistants can be given work at once by asking for both and waiting.
• A task is a single question and a single answer: nothing is queued, nothing runs in the background, and nothing is written into a conversation — so say everything the assistant needs in the task itself.
• An assistant given work this way cannot hand work on again, and a key already narrowed to one map or branch cannot hand out any.
• This is the one thing a connected client does that spends your AI allowance, because the assistant thinks on our side rather than in your client — and a refusal says how much is used and when it resets.
• The list of your assistants now says how many blocks each one looks after, when its branch last changed, and what handing out work has cost this week.
• Seating one is now something you do on a screen: a new API key asks which map and which branch it reaches, and a Telegram bot is seated the same way on its own card.
• A key’s reach is fixed when it is created, so the form says so and every key states its reach in words.
• People on a map now lists the assistants pointed at it, the branch each one has and how much of the map it reads — with no key or token anywhere in it.

A map you make now arranges itself from the first block, so neither you nor an assistant has to decide where anything goes — and the sign-in that gets you here finally has a name.

What changed

• The graph panel no longer flags «loose groups» or «bridge blocks»: the first got worse the more of your map you connected, and the second described every section heading. What is left is the reading that works — pairs of blocks that belong together and are not linked.
• There is now one tool for that reading rather than two. A connected assistant asks for unlinked pairs once, and gets more back: up to twenty instead of three, narrowed to a single block if you want, and the pairs that are close in meaning beside the ones that share neighbours.
• A map can be renamed from an assistant — its name and its one-line description, and nothing else: not the theme, not the layout, and not who can see it.
• Every new map starts as a tree and keeps itself tidy as you connect blocks; maps you already have are untouched.
• An assistant can move a block to where it belongs in one step, instead of unpicking links by hand — and it says which links it removed.
• Reading a map hands back the id of every cross-connection, so a link that is wrong can be fixed from the same read that found it.
• A map read can say when each block was written and when it was last changed — so an assistant picking up an unfinished map can tell what is new.
• An assistant deleting blocks reaches only the map it was working on, and says what it removed.
• A shape a map cannot draw — two blocks each set as the other’s parent, or a link from a block to itself — is refused as you send it, instead of accepted and then quietly hidden.
• Editing the same block twice in one batch is refused rather than done twice and reported wrongly.
• Sending the same map twice — after a connection drops mid-build — is now called out: the reply names the blocks it just made a second copy of, and the tools say how to retry without duplicating anything at all.
• What a map holds now reaches an assistant as text rather than as instructions: each block’s words come back inside a marked-off block, and the reply says plainly that a map’s name, titles and bodies are content, not commands — which matters most for words somebody else wrote onto your map.
• A block’s text comes back exactly as you wrote it; its headings are no longer quietly shifted a level down on the way out.
• A guided interview leaves a trace after it ends: its answers can still be read, and starting a new one on the same map hands them back instead of asking you everything again.
• An interview closed with questions unanswered no longer says it is finished and lists what is still to ask in the same breath.
• A document that asks for nothing is refused with a reason, instead of answered exactly as a successful one was — which is what left an assistant re-sending it.
• When a connection cannot be drawn, the reply names which of the two blocks is the one to fix.
• An id left empty is called a missing value rather than an invalid one, and no error hands back a hundred and twenty characters of pattern any more.
• A complaint about a mistyped field no longer reads as “somebody deleted your block”.
• A block whose title is nothing but spaces is refused, instead of saved and then shown as “(Untitled)”.
• Deleting a block names the children it left hanging, and says they were not re-hung for you.
• An assistant that re-hangs part of your map has to say so, in its own sentence, instead of only mentioning that blocks moved.
• A map that places your blocks for you no longer tells you to space them out better.
• An assistant building a map sends what a block is, and the map decides where it sits.
• It can write a tree by naming each block's parent, with no list of links at all.
• It can add a line to a block — a diary, a log, a running list — without rewriting what is already there.
• Your sign-in is a Corlecti ID, and every screen calls it that.
• The beta waiting list writes to you — when you land on it, and again when a place opens.
• A new account starts on the cheaper of the two models that need no key of yours, and Settings lists it first.
• Arrange, Colour map, Expand all text and Focus are glyphs in the toolbar again.
• Several top-level blocks lay out side by side instead of in a column.
• The Telegram bot's /clear can be confirmed.

The parts of Corlecti Map that run without you — your MCP connector, your scheduled maps, the Telegram bot, the notification digest — could not see which plan you were on, so every one of them treated you as if you were on the free one. Now they know.

What changed

• MCP keys are measured against the plan you actually hold, not against the free ceiling.
• Scheduled maps, the Telegram bot and voice notes all use your real allowance.
• If we cannot reach your account for a moment, nothing refuses you — your work is still counted, and only the 'no' waits.

The plan page now names every limit the product actually enforces, including the two it had never heard of — and Workspace got bigger.

What changed

• Maps on a schedule are on the plan page at last: two at once on Free, ten on Pro, twenty-five on Workspace.
• Export — Markdown, PDF, slides and a full copy — is listed on every card, and it is free on every plan.
• Importing a PDF or a web page is listed too, and it was never a paid feature.
• Workspace: 100 maps, twice the MCP calls, twice the keys.
• The Workspace card no longer promises unlimited tabs on a map — it says thirty, which is what it has always been.
• Settings ▸ Usage shows how many schedules you are using, beside how many you may have.
• An account granted Workspace by hand now gets Workspace everywhere, not only where people meet.

The plans live in the product now. Every refusal leads to a page that knows which plan you are on, what the next one adds, and what it costs.

What changed

• A plan page inside Corlecti Map — the plan you are on is marked, and each card shows only what it adds.
• Monthly or yearly, with the price on the card and the price at the checkout always the same offer.
• Read it without signing in: the pricing page on the site opens it directly.
• “Upgrade plan” stays in your tab now, instead of opening another one.
• The plan you are shown is the one your account can actually use — never one the product would refuse.

Your account has two names now — one for billing, one for how people see you — and your maps rearrange themselves by travelling to their new places.

What changed

• Settings ▸ Account: set a full name for billing and a displayed name for everything else.
• Comments, messages and files say who wrote them by name, not by email address.
• Typing @ in a shared map offers people by name.
• Change the sort order and the cards move to where they now belong.
• Pin a map and its card flies to the front of the shelf.
• Filtering fades out what no longer matches, and the rest close the gap.
• Deleting a map takes its card away; duplicating one puts the copy where it belongs.
• A control that would not actually move anything animates nothing.

A skill can be written for a few maps instead of all of them, and can be handed to the assistant without being asked for.

What changed

• Choose the maps a skill applies to, instead of every map you own.
• Map settings → Skills: which skills this map is built with, and a trash to delete one you wrote.
• Ticking a skill on a map adds that map to its list — the same decision from either end.
• Mark a skill “Always in context” and the assistant is handed it in full every time it opens that map.
• It works the same in the chat panel, in Telegram, on scheduled runs and over MCP.

One theme, several bots, and a way in for people who never had a username.

What changed

• Corlecti Map has one theme, and it is dark.
• Settings → About → Version history: every release since the first day, and which one you are on.
• Connect several Telegram bots, each one pointed at a map you already have.
• Connecting your own bot no longer needs a link — the first message claims the chat.
• Hold the door open for a number of people and a length of time instead of typing usernames.
• Send the bot a voice note and it treats the words as if you had typed them.
• Every “not included in your current plan” now says what to do about it.
• Open the “Thinking…” line and watch what the assistant is doing, and how long the run took.
• A long answer from your bot no longer arrives with a piece missing.

Telegram stopped being one shared bot, and the assistant learned to lay out what it adds.

What changed

• Bring your own Telegram bot — paste a BotFather token and it is yours.
• Your own bot is a chat, not just a capture box.
• Let somebody else write to your bot.
• An MCP key can be bound to a single map.
• The AI allowance is paced by a day and a week — there is no monthly figure to misread.
• A refusal can tell “you used your week” from “your plan shrank”.
• Storage is stated by kind, per account and per map.
• A branch an assistant adds lands under the block it names, and “Keep this map arranged” covers chat and MCP.

The product got its name, its address and its three plans.

What changed

• Every user-facing string says Corlecti Map, and one hostname serves the app and the marketing site.
• Three plans, with numbers on every row, and a voice allowance measured in minutes.
• Read sharing is free on every plan — an editor is what costs.
• “What should I link?” answers with meaning as well as structure, and the graph draws what is near but not linked.
• The block panel has a Nearby section, and one press turns a suggestion into a link.
• The graph lens no longer freezes the tab while it draws.
• A shared map you cannot use is no longer offered, and a reply cannot reach a conversation you were not shown.

Three new ways to put something on a map, and one place that tells you what happened while you were away.

What changed

• Comments — leave a remark anywhere on a map, beside a place and not only on a block; give it a colour, drag it, resolve it.
• Draw something, put it on the map, and open it again later to change a line.
• Say something out loud in the map's conversation and read it back transcribed.
• Record your screen from the map's conversation, with or without your voice.
• Notifications — an inbox with a count beside the three lines, and delivery by email or through the Telegram bot.
• A map can tell you which of your notes belong near a block.

A map stopped being one person's. It grew members, roles, a conversation of its own, and a shelf for what the work produces.

What changed

• Invite people to a map as a viewer, a commenter or an editor — and see who is on it without leaving the map.
• Give somebody one branch of a map to work in, and nothing else.
• A conversation beside the canvas for the people on a map — not the assistant's chat.
• Media and files — everything attached to a map, in one place.
• A strip of the links a map's work depends on.
• People you share a map with can reach it from their own AI tools, behind a switch you own.
• Settings shows what your plan gives you, and how much of it is left.

The instruments needed to run the product, and the last of the AI surface only Corlecti could reach.

What changed

• A deploy-wide analytics page over accounts, maps, usage and sharing.
• MCP has a call quota of its own — counted, never charged.
• Every plan allowance, map slot and byte budget is one row in one table, adjustable without a deploy.
• Duplicate a whole map.
• Write your own skills and scenarios from any MCP client, and run a scenario end to end from one.
• Telegram and scheduled runs draw on your plan's allowance like everything else.
• The assistant asks before it rearranges a map you have already built.

The meter, the limits and the security pass that had to come before anyone else could be let in.

What changed

• Settings → Usage: how much of your plan you have used, which models it went on, and when the window resets.
• An API key per provider, kept at once, instead of one key per account.
• Ten maps on Free, an eleventh available as a slot, and one storage budget shared across every map.
• Overage — choose to keep working after you hit your plan limit.
• The paid parts of the app refuse to be called in a loop.
• Add Corlecti Map to Claude in a browser: the MCP server speaks OAuth.
• A large map can be read in parts, or as a document instead of as data.
• A pass over sign-in, the app's own web address, framing, upload size and the activity feed.
• The dashboard and the chat panel open faster, and stay fast as an account grows.

A way to work on one branch of a large map without the rest of it shouting.

What changed

• Focus mode — click a block and the rest of the map steps back; pin it to keep the focus where you put it.
• Search takes you to the whole branch, not just the block.
• A “+” under the selected block adds the next one, connected and placed, as many times as you like.
• The Arrange menu can keep a map in its shape by itself, re-arranging as you link and delete.
• The assistant asks its questions all at once, in a form you fill in.
• A map opens the way it is built, block by block, instead of blinking into place.
• Scenarios and skills are one press from the chat input.

Four things a map had been missing, and a second assistant for the graph.

What changed

• Ctrl+Z undoes what you just did on a map.
• Colour map — one press gives every branch its own hue, links included, and reshuffles each time.
• Pictures on the map: paste a screenshot or drop a photo, and it comes out in the PDF and the Markdown.
• Talk to the map builder — a microphone beside the send button.
• A second assistant for the graph lens: read-only, proposing links you approve.
• Long builds follow a written plan instead of running out of room.
• Attach a PDF, Markdown or text file straight to a chat message.
• Connection angles became a setting for the whole map.
• The graph view shows your colours and your map's hierarchy.

Two days that rebuilt every surface. The toolbar, every dropdown, the dashboard, the drawer and the block panel became one system instead of a dozen one-offs.

What changed

• The board toolbar is one row of icon buttons, with a keyboard shortcut for everything it switches.
• Every dropdown in the app is the same dropdown — positioned, keyboard-driven, and never off-screen.
• Map presets — one press instead of nine decisions, with your own presets saved beside them.
• The block panel: a block's colours, size and type in one place, over the map.
• The dashboard opens on your maps — pin them, sort them, search them, give each one a colour.
• Sharing is one row: choose who a map is for, and invite someone to edit it, not just view it.
• “Are you sure?” is the app's own dialog now, not the browser's.
• Markdown has a real theme, and the editor is colour-coded as you type.
• Every menu can be driven from the keyboard, and submenus work on touch.

A map could be laid out by its own structure, and remember the shape it was given.

What changed

• Eight shapes in an Arrange menu, laid out from the map's actual structure rather than from geometry.
• The shape is saved on the map, so what an assistant builds over several passes stays consistent.
• Arranging takes a snapshot first, so it can be undone.
• Blocks are given rows tall enough to expand into.
• A block's role can decide its size as well as its colour.
• Arranging a large map is one request, and the view reframes onto the result.

Corlecti Map stopped being something you only use with the tab open.

What changed

• Capture from Telegram — any message becomes a block.
• Scenarios — an interview that asks one question at a time and builds the map from your answers, in the chat or in Telegram.
• Search everything you have written, across every map.
• A third graph lens: close in meaning.
• The graph names where the structure is thin.
• Snapshots and version history in every map's settings.
• Scheduled maps — a map that runs itself on a cadence.
• Write your own skills and scenarios without touching SQL.
• Give the chat a document or a link and get a map back.

A dashboard to start from, four ways out, and a second way to read a map.

What changed

• A dashboard of every map you own, ordered by the map hierarchy.
• Export to PDF — a paginated document with a cover, contents and numbered sections.
• Export to a PowerPoint deck.
• Structured Markdown export, ordered by the map's own arrangement.
• Publish a map to a read-only link anyone can open without an account.
• The graph lens — a second, read-only reading of the same blocks and links.

The map got a settings dialog and, underneath it, the first rules about where things may sit.

What changed

• Map settings — a dialog over the canvas: appearance, layout, statistics and your account in one shell.
• Blocks no longer collide: a layout engine gives every block room and pushes neighbours aside.
• Statistics in two scopes — this map, and your whole account.
• Per-map appearance: canvas backdrop, minimap and accent.
• Maps an assistant builds come out in one colour scheme, coloured by what a block is.
• New blocks appear in reading order instead of all at once.
• Tidy up — pull every block up and close the vertical gaps.

The largest change in the product's life. A map became a document an assistant can write, and Corlecti Map grew a chat, an MCP server and a settings hub.

What changed

• One canonical document for a map, with a transactional import everything else now goes through.
• Copy, cut and paste blocks with the connections between them.
• A chat on every board: describe a map and watch it appear.
• An MCP server — read and build your maps from Claude Code, Claude Desktop or any MCP client, with a personal key you mint yourself.
• A Settings hub, four AI providers, and your own API key encrypted at rest.
• Chat history persists per map, any message can be edited and resent, and a long build streams as several messages.
• Single-entity tools, so the assistant edits one block instead of rebuilding a cluster.
• A collapsed block is a header strip again; an expanded one grows to its own height.

Blocks stopped being one shape.

What changed

• Drag a corner to resize a block; the size presets set width and height together.
• Expand a block to read its notes on the map, and Expand All or Collapse All for the whole board.
• Connections can carry an arrowhead — at the start, the end, both, or neither.

The shared password went away. Corlecti Map became a per-account product, and every row learned who owns it.

What changed

• Corlecti sign-in replaces the single shared password — no separate registration.
• Maps, blocks and connections are scoped to the account that made them.
• Ownership is decided on the server and cannot be sent by a client.
• Someone else's map answers “not found”, never “forbidden” — the app never admits a row exists.
• A portal can no longer link to, or preview, another account's map.

One board became many, and a map could point at another one.

What changed

• More than one map per account, listed as a tree.
• Portal blocks — a round node that opens another map when you click it.
• Sub-maps: a map can belong to a parent map, and the sidebar shows the hierarchy.
• A portal picks its target from a searchable list instead of a pasted id.
• Style inheritance — a child block takes its parent's colours.
• The block types were cut back to the three that earn their place.
• You can zoom out far enough to see a large map whole.

The app took on the Corlecti identity, and learned light and dark.

What changed

• The Corlecti palette — warm neutrals and one muted teal accent; Fraunces, DM Sans and JetBrains Mono.
• A light theme and a dark theme, with a switch in the header.
• Text colour is computed from the colour behind it, so a custom fill stays readable.
• Block presets became theme-aware and invert with the theme.
• The canvas, the minimap and the dot grid follow the theme.
• Markdown export — the first way a map left the app.

The board stopped being one colour, and gained a sidebar to move between boards.

What changed

• A board sidebar — switch boards and edit a board's name and description without leaving the canvas.
• A block toolbar: background colour, border colour and four widths.
• A connection panel: ten colours, solid or dashed, and delete.
• Goal connections — a dashed line for a secondary relationship.
• Select mode — drag a box and move several blocks at once.
• A block can be switched off: dimmed, blurred and not editable.
• Blocks are added from the toolbar, not by double-clicking the page.

The first thing that worked: an infinite board you could put notes on, connect, and come back to.

What changed

• An infinite canvas with blocks you place, drag and connect.
• A block holds a title and Markdown notes.
• Everything persists and survives a reload.
• One shared password, and a deploy that runs itself.