Repository navigation
Releases
OpenCognit 0.5.1 — Cosmos
OpenCognit 0.5.1 — Cosmos.
The graph gets two new views: the task wall traces every task from start to commit, and the Cosmos is now a real 3D scene. Plus Fable 5.2 and the new home at opencognit.com.
New: task wall (Graph → Tasks)
- Every task as a path across cards: origin → agent (with limit failover, the replacement too) → read, edit, tests, guardrails → approval, review, commit, PR, done
- Light threads draw the path step by step; anything waiting for you (approval, question) shows as a warning band
- Data bar with the log, running tasks (pin one to hold it), the quota of your CLIs and the latest guardrail decisions
- Click an agent card to open the chat with that agent
Cosmos in 3D
- An engraved brass disc in space with the gold orb at its center as the light source; planets have a day and a night side and cast shadows
- Planets are areas (folders) on their own orbits, active ones circle faster. Moons are files: dark untouched, blue read, orange changed
- Comets are agents: they orbit the area they work on and arc over to the next one
- Golden threads run from the orb to the agents the orchestrator has assigned; when one finishes, a ring opens
- A change sends a wave to the areas that import the changed file
- A calmer surface: one slim bar with agents, changes today and actions per minute; "More" opens the live ticker, names appear on hover
- Without WebGL, or with "reduce motion" in your system, the Cosmos draws flat as before
Models and usage
- Claude Fable 5.2 in the model list (replacing Fable 5.1)
- Usage dashboard: prices for Fable 5.2 ($8 / $40 per million tokens)
Fixes
- When an agent's turn ends at its usage limit, there's no "… is done" review card anymore — the work is with the replacement or continues after the reset
opencognit.com
- Website, handbook links and the price table now use the new address opencognit.com
OpenCognit 0.5.0 — The Relay
OpenCognit 0.5.0 — The Relay.
No task left behind: when an agent hits its usage limit, another account or another CLI takes over. Plus signed updates, more protection for your credentials, a second style and a new phone app.
Orchestrator: keeps going at the limit
- Limit failover: when an assigned agent hits its usage limit, its task moves with the same number, file scope and progress so far to a free account of the same CLI, otherwise to another signed-in CLI. Accounts that are at their own limit are skipped
- The chat shows who takes over, with a button to open it. The orchestrator is only woken with the replacement's report
- If no replacement is free, the agent waits and continues on its own after the limit, instead of counting as "finished without a report"
- Can be switched off under Settings → System → "Hand tasks over at the limit"
Security
- Signed updates: the app only offers updates signed with OpenCognit's release key. Rejected versions are shown in a notice
- Protecting credentials: when an agent reads credentials or uploads files (curl with a file, scp, rsync, sftp, nc), it needs your approval
- Environment without secrets: agent terminals no longer receive credentials from environment variables. The CLIs stay signed in
- Test guard: if an agent changes or deletes existing tests, the mandatory check says so. Levels: off, report (default), strict. Strict treats it like a failure
- New "Security" tab in the settings collects every protection switch and shows what is switched off
- Handbook: a complete list of everything the app sends over the network
A new phone app
- Tabs: Home, Orchestrator, Agents, Tasks, Inbox. The neurons are one tap away from Home
- Home shows what needs you first: open approvals, a task field for the orchestrator and every running agent
- Orchestrator: pick the workspace at the top, the input always sits at the bottom
- Agent: interrupt via the "…" menu, calmer approval buttons
- Settings on the phone: language, style, computer, version and sign out. Signing out invalidates the pairing on your computer
Look and CLIs
- "Neutral" style: graphite or white with a blue accent, as a second style next to gold. Settings → Appearance → Style
- CLI updates: Settings → System shows the installed and the latest version of every agent CLI. "Update" types the CLI's own update command visibly into a terminal
- Community: Discord, X, website and releases in the "About" tab
Price
- Launch price for the first 50 buyers: €39.33 instead of €69. The discount is already applied in the buy link
OpenCognit 0.4.0
OpenCognit 0.4.0 — The Loom.
Many agents, one fabric: the app now checks for itself what agents report as done, merges their work into your project one piece at a time, and shows all of it in a new neuron view.
Orchestrator: checked, not trusted
- Mandatory check: when an agent reports "done" and the project has tests, OpenCognit runs them itself. Only then does the task count as reported. The orchestrator sees the result in the task line and does not have to check again
- Sent back on red: if the tests fail, the output goes back to the same agent automatically, up to twice. The orchestrator is only woken once the tests are green or the attempts are used up
- Shared folder: if several assigned agents work in the same folder, a failure is not sent back automatically, because it may belong to another agent. The orchestrator is told why
- The mandatory check follows the setting "tests after every turn": if you switched that off, there is no mandatory check either
Merge queue: many worktrees, one project
- Merge queue: the orchestrator merges the worktrees of several agents one after another. Before each one the app saves a checkpoint, afterwards it runs the project's tests
- Red after merging: that one worktree is rolled back, the others continue. If a change no longer fits the project, the project stays untouched
- One confirmation for the whole batch instead of one per worktree. As before, nothing is committed
Load planner
- Test runs started by the app (after an agent's turn, mandatory check, orchestrator check, merge queue) share the machine's slots: one run per four logical cores, only one when memory is tight. The rest wait their turn instead of all starting at once
Neurons: a new look
- Fibres with body and depth of field: branches taper towards the tip, the selected neuron is in focus
- Colour says who, shape says what: pulses run along the fibre in the agent's colour. Reading is a fine pulse; on a change the whole branch lights up to the file
- Afterglow: a touched file glows in the agent's colour for two minutes
- Echo across imports: after a change, a faint pulse runs to the files that import the changed one
- Agents as rings at the cell body, each with a name tag. Waiting agents circle the core as dots; from eleven on, one ring with a counter stands for all of them
- Flight to other projects: click a distant network and the camera glides there, without a loading screen and without a jump
- Smooth on large projects too: with 1,200 files a frame takes about 6 instead of 18 milliseconds
- The orbiting diamonds for memory entries are gone from this view; they remain in the other graph views
Security
- Sharper guard: destructive commands are blocked, deleting outside the project needs your approval
- Single instance per user folder; confirmation on close and on update while agents are working
Bug fixes
- Neurons: the glow of touched files never cooled down, and the file names at the tips never appeared
OpenCognit 0.3.0
OpenCognit 0.3.0 — The Conductor.
The biggest update yet: OpenCognit comes to your phone, works fully local, speaks GitLab — and the orchestrator now reliably leads a dozen agents at once.
OpenCognit on your phone
- Phone app over your own tailnet: Settings → Notifications → "Remote control from phone", scan the QR code — done. No account, no cloud, never exposed on Wi-Fi or the internet (Tailscale only, 256-bit token)
- Control room: who is working, who is waiting, today's API equivalent, all workspaces with their agents
- Agents: follow the terminal live, type, extra keys (Esc, Tab, arrows, Ctrl+C), context and quota at a glance, interrupt an agent
- Terminal at phone width: the terminal resizes to fit your phone — no sideways scrolling; the desktop shows "Open on phone" meanwhile and gets its size back afterwards
- Workspace on the phone: agents, changes (view diff, commit, send back to an agent) and the orchestrator chat — just like the side panel on the desktop
- Start new sessions from your phone — in any workspace, with any configured CLI
- Inbox: approve or reject guardrail requests with one tap, answer agent questions directly
- Neurons and tasks (board, routines, night jobs) on the go
- Everything runs through the same guardrails as on the desktop and lands in the audit log
Orchestrator: reliable even with many agents
- Numbered tasks: every orchestrator task moves through fixed stages (queued → delivered → working → reported). The orchestrator always knows where each agent stands instead of reading screens
- Agents report back themselves: done, blocked or question — with a summary and the changed files. The report reaches the orchestrator in full
- File scopes: each agent gets its own files or folders. Edits outside are rejected, and two open tasks never share a file. Tested with 12 Kimi agents in the same project
- Wake in batches: finished agents get their next task right away instead of waiting for the slowest one
- Survives restarts: open tasks are never lost, queued ones get delivered, every agent resumes its own conversation
Local models
- "Local models only" per workspace: Codex works with Ollama or LM Studio on your machine; cloud CLIs, threads and the orchestrator are locked there — for code that must not leave your computer
- Switch it on in the plus menu or when setting up new sessions, with Ollama status and model picker
- "local" badge in the sidebar
Package protection
- Before an agent installs packages (npm, pnpm, yarn, bun, npx, pip, uv, poetry, pipx), OpenCognit checks for look-alike names of popular packages, made-up packages that don't exist, brand-new packages and install scripts — anything suspicious needs your approval
Hand-off between CLIs
- "Hand off" in every session header: another CLI continues in the same folder — with the task, the Git state and the latest output. Handy when your quota runs out
GitLab
- Merge requests from an agent's worktree and pipeline status as a badge — self-hosted GitLab servers included (via
glab)
Usage
- Local models count as €0 and as "processed locally"
- CSV export per period and project — for billing per client or for your accountant (German Excel format)
Neuron universe
- Zoom out to see all workspaces as living neurons around you; where agents work, impulses travel and the cores fire
Window
- Custom title bar like modern apps: minimize, maximize and close inside the app, drag by the header (snap layouts on Windows). Prefer the system frame?
OPENCOGNIT_SYSTEM_FRAME=1
Settings
- Saves automatically: no more "Save" button — changes apply immediately; invalid input stays marked and is not applied
Handbook
- New pages: remote control from your phone (with workspace view), local models, usage; extended: package protection, hand-off, GitLab, neurons, title bar
Bugfixes
- Frozen terminals fixed: sessions that weren't on screen (other workspace, settings, editor …) stalled after a lot of output and stopped accepting input — especially Kimi Code. They now keep running in the background
- Every agent keeps its own conversation: after a restart, Kimi and Codex resumed "the last conversation in the folder" — with several agents or the orchestrator in one folder, they all ended up in the same conversation (tasks landed at the orchestrator). Now exactly the session's own conversation is resumed; "Restart" keeps it too
- No more orphaned processes: closing or restarting a session left Kimi (which ignores the signal) running in the background, holding its conversation locked ("Session in use by another Kimi Code CLI instance"). The whole process group is now terminated
- Agent command in installed builds: delegation (
$OPENCOGNIT_CLI spawn) pointed nowhere in installed versions. It now works everywhere, without Node.js - No more ghost input: when switching workspace or page, the terminal re-answered old CLI queries and the answers ended up as input in Kimi & co.
- Header in narrow windows: the task field no longer slides under the buttons on the right
- The browser preview shows again which agent is driving it (local dev servers were wrongly treated as artifacts)
- The dock toggle in the header is no longer covered when the dock is open
- CSV files are written in one go (never half-finished)
- The English license message showed a wrong price (now €69)
OpenCognit 0.2.4
OpenCognit 0.2.4 — The Spark.
Neuron Graph
- Workspace fly-in: Click a distant workspace → camera smoothly flies there, no hard reload
- Mini neuron networks: Remote workspaces show a real, deterministic neuron network (seeded by workspace ID)
- Synapse spark: When an agent touches a file, the matching node explodes in golden light — like real neurons firing
- Stronger central core: The golden root node breathes and glows in 4 layers
Browser & Artifact Preview
- Artifact Preview: When an agent writes an
.html,.htmor.svgfile, the browser tab opens automatically and shows the result — no dev server needed - F12 = DevTools: Opens Chromium DevTools directly in the embedded browser (also via toolbar button)
- Loading spinner in the URL bar during navigation
- Focus-select: Clicking the URL field selects the full address (like Chrome/Firefox)
- Placeholder: URL field shows "Enter URL or search…" — DuckDuckGo fallback was already in place
Agent Overview
- Dismiss tasks: Hover an open task → trash icon appears → click removes it instantly
Schedule
- Delete runs: Per-run X button (appears on hover), plus "Clear all" in the section header
- Skipped runs (no own persistence) are kept
Bugfix
- Graph/Chat/Terminal tile was blank in some views (CSS conflict with
position: absolute) — fixed
OpenCognit 0.2.3
OpenCognit 0.2.3 — Fixes
Notifications
- One notification per waiting phase — no more repeats while a waiting agent's screen keeps changing.
- No "waiting for you" flood when OpenCognit starts and restores your sessions.
- No system notifications while OpenCognit is in the foreground; the app shows what's waiting.
"Not responding" only when it's true
- Switching panes or views sends focus and mouse signals to the CLI. These no longer count as input, so idle agents aren't marked as stuck anymore.
Neurons
- The network no longer reloads whenever an agent starts or creates a file; the camera stays where it is.
- Agents carry a name tag with what they're doing (editing / reading a file, working, waiting).
- Zoom like in Cosmos: the scroll wheel zooms towards the cursor, plus zoom in / out / fit buttons.
Guardrails
- The TRUNCATE rule only blocks SQL (TRUNCATE TABLE …), not the word in searches or the Tailwind class
truncate. Your saved rule is updated automatically. - A block notice can be closed and disappears on its own after 20 seconds.
Linux: AppImage or .deb · Windows: .exe
OpenCognit 0.2.2
OpenCognit 0.2.2 — The Night Shift
Agents carry on after their usage limit
- When an agent hits its usage limit, OpenCognit reads the reset time from the message (time, date or "in 2 hours") and tells the agent to carry on shortly after the reset — on its own.
- The banner and the status grid show when it will continue; one click cancels. If you type into the session yourself, OpenCognit steps back.
- Your computer stays awake while a resume is planned. Switch it off under Settings → System.
While you were away
- Come back after a break and a card shows what happened: agents, changed files with +/− lines, tests red/green, limits and resumes, and who is waiting for you.
- Every entry jumps straight to the session or to the before/after view.
Neurons: watch your codebase think
- Your project as a 3D nerve network: folders are neurons, files are branches. Reading glows gold, edits discharge, agents fly through as comets.
- Two agents on the same file or folder show up as an orange ring with a warning.
- New control room view: what needs you (approvals, questions, stuck, limits), folders and agents with open +/− lines, unreviewed changes, loop warning, quota forecast, log, raster and heat map. Search and fly to any file.
Editor and chat
- Editor next to the terminals, per workspace. It follows the agents: changed files open as tabs, changed lines light up, before/after view.
- Cards in the orchestrator chat: files with +/−, preview, accept / improve / discard.
More
- MCP overview: all servers per CLI, copy them to other CLIs without keys.
- Agent family tree: who started whom, including CLI subagents.
- Agents can use the built-in browser preview (Claude, Codex, Kimi, Gemini).
- Kimi Code: quota in the sidebar, usage counted, new models.
- More reliable status: lifecycle hooks of Claude, Kimi and Codex, CLI notifications, "provider overloaded" and "not responding" instead of a wrong "working".
- Squarer controls.
Linux: AppImage or .deb · Windows: .exe
OpenCognit 0.2.1
OpenCognit 0.2.1 — Security update
Protected targets
- Some places almost never have anything to do with an agent's work, but can run code later or expose credentials. Writing there now always needs your approval: git hooks, shell startup files (.bashrc, .zshrc …), autostart and cron jobs, credentials (~/.ssh, ~/.aws, ~/.gitconfig …) and packages installed straight from a URL or someone else's repo.
- This applies to shell commands and to the agents' own file tools (Claude Code Write/Edit, Codex apply_patch, Gemini write_file). Everyday work stays untouched.
Download first, run later
- If an agent downloads a file (curl -o, wget, curl > file) and later runs it in the same session, that now needs approval ("Downloaded script") — even though each step looks harmless on its own. Thanks to the community for pointing this out.
Under the hood
- Electron 44 (actively maintained, fixes known Electron security issues).
- Strict Content Security Policy in the app window and tighter permissions (only microphone for dictation, clipboard and fullscreen).
- Updated test and build tooling.
OpenCognit 0.2.0
OpenCognit 0.2.0 — More agents. More control.
Every agent command checked — now for all five CLIs
- Kimi and Hermes in the terminal now run through the same guardrails as Claude Code, Codex and Gemini: dangerous commands are blocked, risky ones wait for your approval, and secrets are stopped before they get committed.
- Your own Kimi and Hermes configuration stays untouched — OpenCognit gives these sessions a configuration folder of their own, and a renewed login carries over to your real one.
New: Check log
- Settings → Check log shows what the guardrails blocked, held back, approved or denied — with the rule, the command (secrets redacted), where it came from, the agent and the project. Filter by "Blocked" or "Approvals". Stays local.
Orchestrator that reports back
- The orchestrator now gets notified when its agents are done and picks up from there — no more checking in by hand.
Usage and quota
- Usage dashboard in euros across all your CLIs, read from their local logs.
- Quota (5 hours / 7 days) straight from your Claude terminals.
Schedule
- Routines and night jobs in one place. Routines no longer start twice, old runs are cleaned up.
Look and feel
- Status is now a small 3×3 grid everywhere — each state with its own colour and motion.
- Kosmos gets the start screen's background: a warm glow around your project and golden depth particles.
- Opaque notices and test-light output, a tighter quota card, cleaner charts.
Fixes
- The status-line bridge no longer breaks when it runs without a file path.
OpenCognit 0.1.5
OpenCognit 0.1.5
New look for Kosmos
- Kosmos now looks like an astrolabe: a clock dial around your project, an afterglow on the files your agents touched, and a calm "rest" and a lively "work" state.
- When you open the graph, particles fly in and assemble it (a short entrance; switching views uses a shorter one).
- Your project sits in the middle as a golden orb — the same orb as dictation.
- Connections come from imports and from files changed in the same commits. There is one node for the orchestrator.
- Rings view shows your project as a cross-section of the sun.
New: Replay
- A timeline under the graph with one lane per agent: work blocks and commits for today or the whole week.
- Drag the playhead or press play to watch your day again in Kosmos. Click a block to see which files were read or changed.
Chat next to the graph
- Click an agent in Kosmos and its terminal opens on the right; tabs for the orchestrator and every session. You never have to leave the graph.
- From the editor, "Back to graph" takes you straight back.
- Switching workspaces while in the graph keeps you in the graph.
Start screen
- New rotating statue artwork. The start screen stays dark in light mode too.






