Studio Timeline Improved
Improvements to the 88 MCP Desk tools
We audited and improved our desk tools with models with small limits, still receiving a short list with full descriptions. They now also get one extra tool, call_tool, which runs any other tool by name. Its description lists the names that were left out, so the model can always find them.
The 36 tools now improved and fixed: (the tag browser, vaults, canvas boards, bases, templates, plugins and publish).
Two of them, vault create and switch, and “save a note from a template”, need the live memory, so they tell you to use the vault picker or the Notes panel instead of failing.
We checked all 88 with an empty call. The ones that need an argument now say which one, in a sentence, instead of throwing an error from deep inside.
A model that comes with your licence, and a way to top it up
VEKTOR Included gives you GLM-5.3-Flash as a model to use in the Cli/Gui. We changed the allowance: 1 million tokens for the trial week, then 10 million a month, and one Reset Quota press a month that gives another 10 million.
There is now a Top up button added. $5 buys 10 million tokens. Top-ups are spent after your month runs out, oldest first, and each one lasts two months from the day you buy it.
A new card at the top of Config shows what is true right now: your plan and when it expires, this month’s use and when it resets, and each top-up with what is left and the day it ends.
Memory you can check
Most memory tools tell you what they remember. Few show their working. We updated this into four parts for version 2.0.5.
Receipts. Every recall can leave a receipt: which memories fed the answer, in what order, and where each came from (you, derived, the agent or an outside source). You can reaffirm one, replace it with a correction, or dismiss it, and a corrected memory stops coming back.
A Review page. Open /review for one list of everything waiting on you: Faraday holds, half-finished records found at start-up, flagged outside content and proposed routines. Only a person can decide, and anything that came from outside content can never be approved “for good”.
Memory that self-improves. During sleep, groups of three or more related notes can be summed up as one belief that links back to its sources and carries a confidence score.
The sources are archived only after a week, never in the middle of a session, and you can revoke a belief to bring them back. On a synthetic 180-day test, the share of duplicate slots in the top five recalls fell from 65% to 8% and the tokens per distinct fact fell from 23.6 to 9.0, with the same recall.
A signed history. Memory writes and removals can be recorded in a log where each entry is signed and linked to the one before. Edit a row, delete one or swap in another key, and the check fails. A removal keeps the hash of what was deleted and the reason, but not the text.
Studio timeline
Three changes this round, all tested.
Freehand masks. Mask mode has a Polygon button. Drag a corner to reshape, drag the + on an edge to add a corner, double-click a corner to remove it. A clip can also carry several masks together.
Smooth motion paths. Press Smooth path and the straight lines between position keys become a curve through the same points, with no jolt at each key.
Reversed clips. Trimming, splitting and joining a clip that plays backwards used to cut the wrong end of the footage. The picture at the cut now stays where it was.
Smaller things you will notice
Workflow templates open as a form: Cover Letter Writer has its own boxes for job title, company and tone instead of one blank text field.
Pixabay videos show their thumbnails again (Pixabay moved where it keeps them). Enlarging a picture in Sample Media goes through the app, so sites that refuse direct links no longer break it, and large local pictures open up to 120 MB.
Compress now sits right under Convert in the Inspector as a combined panel.
Try it
Update with npm install -g ./vektor-slipstream-2.0.5.tgz and open the app. Config now starts with the usage card. Visit /review on the graph server to see the inbox. vektor doctor --timing measures how fast your computer finds a memory, using queries from your own memories, and sends nothing anywhere. Consolidation and reflex suggestions are off by default; the changelog says how to turn them on.
Everything else in 2.0.5
New
New: freehand masks on the timeline. A Polygon button in Mask mode makes a shape you can reshape: drag a corner, drag the + on an edge to add one, double-click a corner to remove it. A clip can also carry several masks together.
New: smooth motion paths. A Smooth path button turns the straight lines between position keys into a smooth curve through the same points, with no jolt at each key.
New: Usage and plan in Config. A card at the top of Config shows real figures from the VEKTOR server: your plan and when it expires, this month's allowance used and left, when it resets, when Reset Quota is available, and each top-up with how much is left and the day it expires, plus a Top up button. Refresh updates it.
New: vektor doctor --timing. Measures how fast this computer finds a memory (median, 95th percentile and slowest, on queries taken from your own memories) and says whether it is within the 50 ms target. It runs on this computer only; nothing is sent anywhere.
New: top up with a card. When you want more of VEKTOR Included, a Top up button on the Included card buys 10 million more tokens for $5 on a secure Polar page; your card is never typed into the app. The tokens are used after your month's 10 million, and each top-up lasts twelve months from the day you buy it. The card shows how much top-up credit is left and when the oldest expires.
New: two more free NVIDIA models. Mistral Large 4 (1T-total, 52B-active, text and image) and Nemotron 3.5 Bolt E4B (a small, fast model for reasoning and tool calling) are in the NVIDIA list. Both answered a live test, Bolt in under a second.
New: Review page. Open /review in the graph server for the HOLD inbox (approve once, approve and remember, deny), a "Why?" view that shows a recall's receipt with Reaffirm, Correct and Dismiss on each memory, and the profile with its version history and rollback. Local only, same guard as the other panels.
New: Windows key custody for the signed log. It protects against copying the key to another account or machine; it does not protect against malware running as you.
New: action traces for reflexes (opt-in). Arguments named like keys, tokens or passwords, and any value that looks like a key, are replaced by "[redacted]" before saving. Off by default.
New: Why did you say that? (recall receipts). Every recall can leave a receipt listing the memories that fed the answer, with their rank, score and where each came from. A memory can be reaffirmed, replaced by a correction, or dismissed; a corrected memory stops appearing in recall. Memories now carry an origin (you, derived, agent, external); weighting recall by origin is built but off.
New: HOLD inbox. One list for Faraday holds, audit gaps found at start-up, tainted-content flags, derived beliefs and reflexes waiting for approval. Only a person can decide; tainted items can never be remembered; "approve and remember" is scoped and expires after 7 days at most.
New: memory that forgets on purpose (consolidation). During sleep, groups of 3 or more related notes can be summed up into one derived belief that links back to its sources, has a confidence, and is marked tainted until you promote it. Source notes are archived only after a 7-day grace period and never mid-session, and a belief can be revoked to restore them. On a synthetic 180-day test the store shrank 43% (1260 to 716 nodes), duplicate slots in the top 5 fell from 65% to 8%, and tokens per distinct fact fell 62%, with the same recall.
New: signed, tamper-evident history. An append-only log of memory writes and removals, hash-chained and signed with Ed25519 (built-in crypto, no new packages). Edits, deleted rows and foreign signatures are detected. A removal keeps the content hash and the reason but drops the text.
New: sync without a cloud (building block). Each device keeps its own signed chain of memory events; devices swap what the other is missing, verify it, and rebuild the same state with fixed conflict rules (your edits beat the agent's; a removal beats a simultaneous edit unless you re-add it). Includes rewind, diff, restore and branches ("what did the agent believe on October 3"), and refuses to compare vectors from different embedding models.
New: import and export. Reads ChatGPT and Claude exports and a folder of markdown notes (preview first, resumable, no duplicates on re-run, who said it sets the origin). Exports to an open JSONL format and a markdown vault.
New: versioned profile. Profile facts can be proposed, accepted, pinned or retired. Automatic processes may only propose; pinned facts cannot be changed by them; every change needs a reason and is kept in a history you can diff and roll back.
New: two-way panels and goals as patches. Panels can be listed, read and written by the agent through the gate, with a revision check so an agent write can never overwrite a newer edit of yours. Your edits are recorded at the highest trust. Goals are a list of changes with reasons; the agent cannot change a goal you set.
New: gated connectors. Anything fetched from the web or a calendar feed comes back marked tainted, with hidden text (comments, hidden elements, zero-width characters) removed. After tainted content enters a session, any action with a side effect is held for you. In a 50-payload test, 50 of 50 actions were held and none of the hidden text got through.
New: tools work on models without tool calling. A capability table (native, constrained, text, none) with a 14-day cache, a text-protocol fallback that recovered 13 of 13 printed tool-call shapes with no false positives, a shortlist that sends 9 tool descriptions instead of 118, and a detector for "described the call instead of making it".
New: gated reflexes (opt-in building block). Repeated successful action sequences (5 or more clean runs; any failure or correction rules them out) can be proposed as a template with parameters limited to the values seen. A reflex needs your approval, runs every step through the gate, checks every result, stops and hands back to the model on any surprise, and never runs under tainted content.
New: poisoning red-team pack, markdown projection and intake router. Eight memory attacks run against these modules, 8 of 8 resisted. Memories can be written out as a folder of notes you or another tool can read; your edits come back as your own, highest-trust changes. A rule-based router files each new note as a decision, task, event, profile fact, preference, fact or drop (15 of 15 on a small test set).
New: Desk can check its own health and time its models. Say "health check" or "run a speed test" and Desk reads the app's database, providers and Ollama state, or times a short answer from the local model and each connected provider (time to first token, tokens per second, total time).
New: Studio timeline: on-picture editing. With a clip selected, Transform puts handles on the preview to move, scale and rotate it (Shift snaps the turn to 15 degrees). Mask draws a rectangle or ellipse mask with a Feather slider and Invert. Pin drags the four corners of the picture for a perspective look. A motion path shows the clip's position keys as a dashed line you can drag, and Key position here adds a key at the playhead. Every drag is one undo step.
New: Studio timeline: more clip tools. Play backwards (video clips; the sound is muted), a batch panel that edits scale, position, turn, opacity, speed, fades and blend on all selected clips at once, a Markers list (rename, recolour, jump to, delete), a Compare wipe between the picture with and without effects, Copy keys and Paste keys in the keyframe pane, and Alt+Left and Alt+Right to jump between keyframes.
New: rebindable timeline keys. The Keys button lists every shortcut; Set or Add a new key for any of them, Reset one or all. Q and W trim to the playhead, comma and full stop nudge a clip, and square brackets step between clips. Choices are kept in this browser.
New in the timeline: Find scenes, Auto reframe, Stabilize, Sync clips and Multicam buttons. Find scenes drops a marker at each cut of the selected video. Auto reframe crops a wide video to a tall project (or the other way) so the frame follows the action. Stabilize counter-moves a shaky shot. Sync clips lines cameras up by their sound, and Multicam cuts a new track between overlapping clips to whoever is loudest. Each is one undo.
New: Audio FX. An equaliser (bass, mid, treble), a compressor and a limiter for the whole mix, in the View group. The export includes them.
New: track sync lock. Press Sync on two or more tracks; a ripple delete or ripple trim on one then also moves the later clips on the others, as long as nothing would be cut or overlap.
Added: more timeline editing. Ten more blend modes with a Blend picker, coloured markers (right-click a marker to change its colour), a volume slider on every track, and Insert copy / Overwrite copy buttons that put the selected clip in at the playhead.
Added: more timeline looks and tools. Diamond, Blinds and Barn door transitions, Ripple and Hue shift effects, plus the maths behind auto-reframe (a crop that follows the action) and scene finding (cuts to markers), tested and ready to wire in.
New: Studio timeline: looks, layouts and text animations. Copy one clip's look (effects, blend, opacity, flip, crop, mask) and paste it on others; arrange clips side by side, stacked, in a grid or picture in picture; make a picture a corner watermark; add a lower third (name and title bar); choose a fade shape; and bring titles in with fade, slide, pop or drop.
New: Studio timeline: TTML captions. Export captions as a TTML file and import TTML next to SRT and VTT.
New: Studio timeline: EDL export and import. Save the timeline as a CMX 3600 EDL (drop-frame timecode, dissolves, speed changes and freezes, markers) and open an EDL as a timeline.
New: flip, crop, solo and track colours in the Studio timeline. Mirror a clip left-right or up-down and hide its edges in the new Frame section, solo a track to see or hear only that one, and give tracks a colour to find them fast.
New: a suggested reply sends with Enter. When Desk offers a next message in the empty box, Enter or the arrow sends it, the way Claude does. Tab still fills it in so you can edit first.
New: Studio timeline: new edit tools. Duplicate a clip (Ctrl+D), close gaps on a track, join two halves of a cut, slip the source under a clip, slide a clip between its neighbours, roll the cut between two clips, and stretch a clip to a new length with its speed adjusting to fit.
New: Studio timeline: motion and ducking. Add Drift, Sway, Breath, Spin or Shake to a clip; it is baked into keyframes you can edit. Duck a music track under voice so it dips while someone speaks.
New: Studio timeline: keyboard editing. Q and W cut a clip back to the playhead (Shift also closes the gap), comma and full stop nudge a clip a frame, and the square brackets step to the previous or next clip.
New: Studio timeline: copy, paste, lift and extract. Copy, cut and paste clips with Ctrl+C, X and V (they paste at the playhead, keeping their layout across tracks), and lift or extract the marked in and out range from every unlocked track.
Better
Better: a tool called without a required detail answers in a sentence. Instead of an error from deep inside the tool (for example cloak_fetch with no link), the tool says what it needs.
Better: workflow templates open as a real form. Running Cover Letter Writer, Ad Copy Generator or any of the 15 templates now shows a card with a labelled box for each detail (Job title, Company, Tone as a list, number boxes, tall boxes for code or work history) instead of one blank text box. Auto-fill with AI fills the boxes, and what you leave blank is simply left out.
Better: Compress sits right under Convert in the Inspector. The list of what can be converted and the installed helper tools moved to the bottom of the page.
Faster: repeated recall. Recall receipts add no measurable time.
Better: Convert and Compress are one page in the Inspector. Both sit under one tab, "Convert and compress", with Compress below Convert.
Better: the buttons in Health are in the same place in every card. Back up now and the two exports, and Run now for memory upkeep, sit at the top right, like Run now for memory consolidation.
Better: the Inspector Compress settings and the Compress button sit on one line. The settings are narrower so they fit; on a narrow window the line wraps.
Better: the Desk toolbar has panel tools instead of formatting buttons. The bold, italic, code, quote, heading, list and link buttons and the citation style menu are gone; Sandbox, Terminal, Files, Flowire and Command line sit in their place.
Faster: film strips on the timeline are remembered. The thumbnails inside a video clip are kept on this computer, so a reload no longer decodes the video again; one strip is made at a time.
Faster: sound waves on the timeline are remembered. The wave drawn inside a sound or video clip is kept on this computer next to the film strips, so a reload no longer decodes the whole audio file again.
Better: each keyframe has its own curve. Click a key in the Motion graph to select it; its ease (including a custom bezier curve) then shapes only the stretch up to the next key.
Better: the Inspector uses your theme colour. The selected tab, the Compress, Convert and Inspect buttons and the quality slider follow the theme instead of plain grey.
Better: every card has its close button on the right and its action buttons in the middle. Look, Show as text, Listen, Copy as Markdown, Save, Email and Print / PDF now line up the same way on all cards.
Better: broken pages are caught before they open. A cut-off HTML file, an unclosed script or a JavaScript syntax error is reported back to the model so it can fix the page, instead of opening a blank tab.
Better: reasoning models get more room to answer. The minimum token budget rises from 2,048 to 4,096, and more reasoning models are recognised, so a model that thinks first has space left for the answer.
Better: your model choice is kept. Desk now answers with the model you picked. Nothing swaps it for a cheaper or faster one, and a second provider is no longer raced against it. Another provider steps in only when yours has no working key or stops answering, and the answer says so.
Better: Animate puts the picture first. The picture area is large on the right and the controls sit together on the left, so a photo gets room to show.
Better: diagrams stay a sensible size. A diagram in an answer is a standard card, and Expand (or a click) opens it large.
Better: Studio Explore uses the full width. of the page, lined up with the Shelf bar above it.
Better: the key status panel is wider, with Test my keys, Find best models and Add or change keys in one row.
Better: JOT card buttons sit in one even row. (Save, Plan, Improve, Summarize, Extract Actions, Save to Memory), all the same size.
Better: model cards are half the width, so twice as many fit on a row in Config.
Fixed
Fixed: trimming, splitting and joining a reversed clip. The cut now takes the right end of the footage, so the picture at the cut stays where it was.
Fixed: 36 Desk tools answered "Unknown tool". The tag browser, vaults, canvas boards, bases, templates, plugins and publish tools were offered to the model but never given the database they need, so every call failed. They now run (vault create and switch and "save a note from a template" say plainly to use the vault picker or the Notes panel, because they need the live memory). All 88 Desk tools were checked with an empty call.
Fixed: browser identities. cloak_identity_create passed its whole input as the name and cloak_identity_use called a function that did not exist. Both work now: create saves a named identity, use loads it (and fetches a page if you give a link).
Fixed: the Top up button could stay hidden after updating. The Included bar's script is now reloaded with the new version.
Fixed: Desk could not reach most of its 88 tools on free and small models. To keep requests small, those models were shown about 20 tools, and Desk then said it had no tool such as cloak_fetch. Every model now also gets one extra tool, call_tool, that runs any other tool by name; its description lists the names that were left out. A tool you name is still added to the short list directly. Checked live: a model ran a tool that was not on its short list through call_tool and returned the real result.
Fixed: Pixabay videos showed empty boxes. Pixabay changed where it keeps a video's picture; the thumbnails now come from its new field, so the video results show again.
Fixed: enlarging a picture in Sample Media could show a broken image. Pictures from Pixabay, Pexels and the other sites are now fetched through the app (the sites refuse direct links from another page), pictures on this computer can be up to 120 MB when enlarged (25 MB when added to a prompt), and if the big version still cannot load, the small picture is shown with a short reason.
Fixed: Desk sometimes answered with its own answer rules. After a web search a weaker model printed the rules block back as its whole reply. Desk now cuts the rules out of any reply; if nothing else is left it shows what the tool returned, or says the model repeated its instructions and offers a retry.
Fixed: web search in Desk and JOT works with any search key. It now tries Serper first and then each other service in turn, and says which one answered. Desk's own agent search also uses every provider and the keys saved in Preferences, not just two environment variables.
Fixed: Desk could not use many of its own tools. Asking Desk to fetch a page with cloak_fetch got "I don't have that tool". A question that looked like plain chat skipped the tools completely (now a web address, a tool name, "mcp", a search service, or a speed or health question keeps them), and 36 of the 88 tools were never offered to Desk at all (the note, vault, canvas, bases and template tools, and the note transforms). Four more were offered but said "not available" when called: the timeline and date tools and the two TurboQuant tools. All 88 are now offered and run.
Fixed: Desk said it had no search tool. Desk can now search the web with a new web_search tool.
Fixed: Desk said it had no Serper or web search tool. When you ask about searching, Serper, Google, Tavily, Brave or the web, the search tools (cloak_search, cloak_search_deep, cloak_fetch) are now offered to the model, also on the shorter tool list used by smaller and fallback models.
Fixed: the Bauhaus controls in Art did nothing on other designs. They only reach the Bauhaus designs and the line overlays, so they now grey out, with a note, when another design is chosen.
Fixed: Ledger buttons ran past the edge of the Records card. Each record now fits the card at any width, and Mark paid, Exclude, Rate and Delete stay on one line.
Fixed: the model list in the + menu of the Coder (and Desk) now opens in full. It was cut off to a thin strip; the list now opens inside the menu, with its search box, and the menu scrolls.
Fixed: Studio pop-up cards now follow the theme. The Audio FX, Markers, Keys and ease-curve cards and the snip editor use the theme's card colours, and their sliders, tick boxes and buttons match it in every light and dark theme.
Fixed: chat search only worked with Serper. The search tool in Desk and JOT chat (cloak_search, and the deep search built on it) now uses every search service you have a key for: Keenable, Serper, Ceramic, SerpApi, Brave, Tavily, Firecrawl, and keyless DuckDuckGo as a last resort, with each one's daily limit. With no key at all, the error says which keys to add.
Fixed: Save picture on the code animation did nothing for some animations. It now saves WebGL animations (they came out blank), animations with no canvas (drawn from the page), and says what went wrong or that the preview did not answer instead of staying silent.
Fixed: the chat said it had no web fetch tool. With many tools connected, the short list sent to the model could leave out the one you named or the web tools. A tool you name (like cloak_fetch), and the web tools when you give a site address or ask about a website, are now always included.
Fixed: "Open it in the sandbox" now opens HTML pages. Run in sandbox opens a .html or .htm file in your default browser. Before, only Python, JavaScript, PowerShell, batch and shell files could run.
Fixed: tool calls a model typed as text now run. When a fallback model wrote a tool call as plain text, such as run_verified(...) or a write_file block, nothing happened and Coder stopped as if the model had quit. Desk and Coder now turn those into real tool calls.
Fixed: expanding a diagram showed a tiny white box. The large view now always sets an explicit size.
Fixed: 3D pages in the Code panel stayed black. A page that listens to the camera controls (for example a physics simulation that redraws on "change") stopped with "controls.addEventListener is not a function". The built-in orbit controls now accept listeners and fire start, change and end, keep a camera the page has moved, follow a changed target, and honour auto-rotate, enabled and the distance and angle limits.
Fixed: a flowchart written on one line now draws. Some models write a whole card on a single line; it used to show as plain text and now draws as the diagram it is.
Fixed: sequence, class and state diagrams draw instead of showing as text. Desk now asks models to write those as Mermaid, which draws as a diagram card.
Run your own race, keep your data sovereign.
Version 2.0.5 is live in the downloads page: https:www.vekmem.com
Full list of changes: https://vektormemory.com/docs/changelog
Novel Technology · Privacy Enhanced · Data Sovereignty · Vektor Memory
600+ models BYOK · 60+ MCP tools · 16 Model Providers

Top comments (0)