Command Center
Direct missions. Watch a live terminal fleet do the work. A knowledge graph and memory that persists. On your machine, under your command.
Runs locally · Bring your own AI subscription · Your data stays yours · Flat rate, no usage billing
Direct. Dispatch. Watch. Complete.
- Windows
- macOS
- Claude (Claude Code)
- OpenAI Codex
Missions, managed by agents
Move a mission to Active and choose to staff it with a fleet: a workspace appears, a commander agent decomposes the goal into operations and fills them with worker sessions. Everything stays visible and steerable, with success conditions as the checklist that gates "done".
- A live mission board (Planned, Active, Completed, Cancelled) with a real-time strip of the agents working each mission
- Missions break into operations, each with its own written assignment, its own agent, its own status. Every mission is a readable file with a full timeline; nothing lives only in a chat scrollback
- The commander coordinates; you decide: message the lead straight from the board, take over any pane, and re-steer the fleet at any moment
Missions → operations → worker sessions · success conditions gate completion
One working loop that you command: you set the goal, agents work in visible terminals, you see everything, and a mission only counts as done when its success conditions are met.
Direct
You describe the idea, the goals, or the mission. An agent researches, plans, and proposes an end-to-end, thought-through mission with explicit success conditions.
Dispatch
A commander agent staffs the mission: worker sessions spin up in real terminals, each grounded with curated context.
Watch
Every agent session is a real, persistent terminal pane, not a simulated readout. Read along, take over any time, or pause the whole fleet with one click.
Complete
Success conditions are checked off one by one. A mission only completes when its checklist is met. You end with verified, finished work, not a chat to re-read.
Agents that keep working, sessions that never lose the thread
Every agent in the fleet runs as a genuine, persistent Claude Code session, pre-grounded with your context. It survives reloads and pop-out windows, and resumes exactly where it left off after you quit and relaunch, the same session, not a fresh one. Because it's real, not simulated, nothing is hidden: read any line, take over any pane, or freeze the whole fleet with one click.
- Persistent Claude Code sessions, pre-grounded with your curated context, pointed at any folder or drive
- Sessions survive reloads and pop-out windows, and conversations resume where they left off after you quit and relaunch
- One click pauses the whole fleet; every control action is a line in an audit log on your disk, next to the full session transcripts
Pause the whole fleet in one click · every action audited · sessions resume on restart
Dictation that never leaves your machine — included in Pro
Hold one hotkey, speak, and an on-device speech model turns your voice into text, typed straight into whatever field has focus, in any app. It runs entirely on your machine: no cloud, no account, nothing sent anywhere. It is part of the base subscription — you do not need Ultra for it.
- Included in Pro and in Ultra: dictation is part of the base subscription
- Runs on your machine: an on-device speech model, nothing sent to the cloud
- Fully offline: dictate on a plane, in a tunnel, with the network off
- No account and no per-word cost: it costs nothing extra to use
- Types system-wide into any app: your editor, your browser, your chat box
- A push-to-talk hotkey you set: hold to talk, or tap to toggle hands-free
Included in Pro · on your device · fully offline · types into any app
Ultra exclusive
The command in Command Center
Summon a voice assistant from anywhere and run your platform hands-free. It reads your project, remembers across sessions, and acts, opening views, arranging windows across your monitors, and driving your terminal fleet through a tool reach you curate: read-broad, mutation-narrow, destructive actions denied, every action pausable and logged. Talk to Grok or ChatGPT: your provider, your account.
- An always-on-top voice widget with a summon hotkey: talk while any app is focused
- It cites its sources: answers grounded in your files, your memory, and your knowledge graph
- Outbound always waits: your voice can compose the newsletter; only your click sends it
Summon from anywhere · hands-free control · a tool reach you can pause
Your entire knowledge as a walkable map
Cortex mirrors folders, documents, and extracted relationships as a real, browsable graph, with an integrated graph database that is ready the moment the installer finishes. The default path costs zero external model tokens; semantic vector search stays a deliberate opt-in with your own key.
- One click, ready: the graph database ships with the installer and is managed for you. No Docker, no setup
- Your agents read the whole project as one compact map instead of re-discovering it every session
- Three depths: structure (free), relationships (no external models), vectors (opt-in), plus zero-AI graph analytics and portable JSON/GraphML export
One compact map of your whole project · 0 external model cost by default
A memory that never forgets, a context that never bloats
Context engineering is the discipline of giving an AI system exactly the right information at the right moment. Here it is built in, not bolted on: six deliberately separated memory layers with written ownership rules, from session context through a searchable cross-agent knowledge base to the permanent strategic board. What one agent learns, every future agent finds; decisions survive restarts, weeks, and context switches. And the pull is engineered: agents orient on the whole knowledge graph for a few hundred tokens, retrieve budgeted answers instead of raw files, and a relevance gate returns nothing rather than noise. The result is not a prompt, it is a system: consistent, reproducible, reliable enough to run without you watching, and fully inspectable the moment you choose to look.
- Hybrid search: keyword and meaning combined, with a relevance gate instead of noise
- Facts evolve with lineage and stay curatable: update, supersede with history, delete
- Curated context packages pre-ground every new session: no re-discovery, no re-explaining
- Every database is visible, backed up, and restorable in the app. Your memory is an asset, not a black box
6 layers · whole-graph orientation in a few hundred tokens · supersede with lineage
Harvest any source, on autopilot
Point it at any feeds, X accounts, YouTube channels, or your own newsletter inbox. It fetches what's new, strips it to clean text, drops duplicates, and files everything on your machine. No API keys, no model cost. It ships with 100+ curated sources: keep them, curate them, or replace them with your own. Video gets its own deep treatment: paste a URL, or schedule a whole channel, and get a report you'd never take by hand — an 8-section analysis with every slide, chart, and number pulled straight off the screen.
- Five harvesting methods, no API keys required: RSS/Atom feeds, X accounts, YouTube channels, email newsletters, full video
- Three-layer duplicate detection: the same story arriving from three feeds is stored once
- Your schedule, from every 15 minutes to daily. It harvests unattended and catches up after downtime
- Video deep-dive: an 8-section analysis per video — the core idea, timestamped takeaways, every tool and framework extracted, a minute-by-minute walkthrough — with frame references anchored to the second
- Content-aware capture: frames are taken where the screen actually changes (slides, charts, code, on-screen links)
- Batch a channel or a conference and get a cross-video trend analysis on top
Any source · zero model cost to harvest · video gets an 8-section deep-dive, timestamped to the second
Put any agent on a schedule
Between harvesting and finished intelligence sits the engine that runs the pipeline: put any agent on a schedule and it works on its own. Nightly research, a morning briefing, a weekly roundup. Every automation is a plain, readable file, not an opaque job you can't inspect. Change the schedule, the model, or the instructions in place; it retries on failure, steps back when a source is misbehaving, and writes every run to a log you can open.
- Every automation is a readable, editable file: adjust the cadence, the model, or the instructions and it takes effect without a restart
- Any cadence, from every few minutes to weekly, or on demand. Watch a run stream live, stop it mid-flight, and read the log of every past run
- It retries on failure and disables itself after repeated errors, so a broken job never runs unattended forever
Schedule any agent · live run output · retries and auto-disable · every run logged
The real routines cockpit, running on a sample schedule. Nothing here is a screenshot.
A calendar that dispatches agents, not reminders
The platform's only date-addressed scheduler. Where automation answers "every morning", the calendar answers "at this exact moment" — and a due entry doesn't fire a script, it dispatches an agent. Agent tasks, whole terminal workspaces, mission fleets, one-off routine fires, scheduled email and plain notes all live on one week grid you drag and drop. Every dispatched entry carries how far it may reach and how you watch it run.
- Six things to schedule on one grid: an agent task, a pre-configured terminal workspace, work on a mission, a one-off routine fire, an email, or a plain note
- A background run sets its own reach, report only, may edit files, or may act outward. A run you can watch and take over in a terminal comes up at full reach, because you are there
- Mark out time and write what should happen in plain language, and an agent turns it into real entries, checking what is already booked first. One action undoes everything it made
- One armed/paused switch is a real kill switch: stop every dispatch without touching a config file, with each firing on an audit log
The only date-addressed scheduler · a due entry dispatches an agent · one switch pauses them all
The real calendar, running on a sample week. Every entry here would dispatch an agent — nothing here is a screenshot.
Subscription & pricing
Not open yet Purchases are not open yet: we are still testing the live checkout, delivery and licensing end to end. Join the list and we will let you know the moment it opens.
Pro
The full command center for one operator
Included
- Mission System: agent driven, start to finish
- Terminal Workspaces: agents that coordinate directly
- Calendar: schedule agents, missions and workspaces
- Email: an agent driven automation system
- Routines: recurring agent work on your schedule
- Observer: harvest X, videos and websites mechanically
- Intelligence: process that harvest into research and news
- Cortex: knowledge graph, knowledge base, artifacts
- Memory: everything it learns, instantly findable
- Agents, skills, protocols: yours to read and edit
- STT: Speech to Text, local and offline
Ultra
The full platform, the production pipelines, and Voice
Included
- Everything in Pro
- Voice: run most of the platform hands-free; its conversational AI runs on your own xAI or OpenAI account
- AI Video Production: from idea through script and storyboard to the finished render
- Atlas Engine: build knowledge products you can sell
- Knowledge library: every product in our catalogue
- Cancel monthly
- Flat rate, no usage billing
- 30-day money-back on annual plans
- Your data stays yours
- Offline-tolerant: a network blip never locks you out
Requirements: Windows 10/11 or macOS · Your own AI subscription: Claude (Claude Code) or OpenAI Codex · Node.js and Python: installed automatically if missing · Optional API keys for extra capabilities (e.g. semantic search, image generation)
2 machines per license
Your starting rate stays locked for as long as you stay subscribed.
In accordance with § 19 UStG, no VAT is charged.
An intelligence desk that runs itself, on your sources
The harvested signal becomes finished, on-brand content: impact-ranked briefings, a live breaking ticker, deep reports, weekly roundups, and bilingual newsletters, all written in your voice, in your languages, published straight to your site. The one step that reaches other people, sending, waits for your click.
- Six finished formats in your voice: briefings, breaking items, deep reports, roundups, newsletters, translations, plus generated hero images
- Impact-ranked by evidence: the story three sources report independently is flagged as the big one
- Seed your brand once and every writer produces in your voice, and every published piece carries an automatic AI disclosure
5 harvest methods · 6 content formats · your sources, your site, your voice
Turn your research into a product you can sell
Point the pipeline at a pile of source material and it builds a finished, two-part product: a structured knowledge base an AI agent consumes, and a human manual filled with ready-to-use prompt blueprints. Each stage is a dedicated specialist — one reads and maps your sources, one researches each section, one drafts, one refines, one finalizes in a neutral voice, one writes the human-facing companion. You watch it advance section by section, approve the scope, and export a cover and a ship-ready bundle at the end.
- From curated sources to a shippable deliverable: seed, research, scope, refine, finalize, and package — each a stage you can inspect and re-run
- Two components ship together: an agent-facing knowledge base and a human manual of prompt blueprints, plus a generated cover
- You hold the gate: approve the scope, steer any section, and export a single ship-ready bundle when it's ready
Curated sources in · a two-component product out · every stage inspectable · a ship-ready bundle
An inbox with an agent that drafts, and a human who sends
Connect any mailbox and the Command Center gives you a cockpit, read, reply, forward, triage, plus an email agent that reads new mail, pulls context from your knowledge base, and drafts a grounded reply. Nothing goes out on its own: every draft waits in a review queue for your approval, unless you deliberately arm auto-send for a narrow, confident lane.
- A full inbox cockpit over any IMAP/SMTP mailbox: read, reply, reply-all, forward, move, and triage across accounts
- An email agent drafts grounded replies from your knowledge base; a human approves and sends, drafts stay drafts until you say so
- An always-on monitor watches for new mail, and past agent runs are kept as readable reports
Any mailbox · AI drafts, a human sends · HITL by default · nothing auto-sent unless you arm it
What you can run on day one
Individually, these are features. Together, they are operations you direct by voice, watch in real terminals, and steer at any moment, the part no single tool does.
A written goal becomes a visible team
Write a mission with a checklist of success conditions and staff it with a fleet: a commander agent breaks it into operations and runs worker sessions you watch in real terminals, reporting back when the conditions are met.
Command the platform by voice
Summon the assistant from anywhere: open views, arrange windows across your monitors, drive and read terminal sessions, walk your knowledge graph. Every action audited, one click from a stop.
An intelligence desk on your schedule
You set the cadence: a breaking ticker every half hour, briefings daily, roundups weekly. Sources are harvested and deduplicated continuously, finished drafts land in your review queue, and one click publishes. Nothing goes out without it.
Knowledge that compounds
Package context once and every agent session you launch starts oriented; what one agent learns lands in a shared memory the next one finds. The system gets sharper the longer you run it.
The platform
The platform underneath
Agent contracts
47+ pre-built specialists covering research, audit, review, engineering, writing, illustration, and end-to-end content pipelines, each a readable, editable file. Adding a new agent means writing one markdown file: clone a proven one, adapt it to your own workflow, and it's live on the next run.
The MCP server fleet
Memory, knowledge graph, terminals, window control, harvesting, email, calendar and mission dispatch, image and video pipelines: exposed as typed tools any Claude session can call, wired from one settings file you can extend with your own.
Skills & protocols
23 skills and a dozen written protocols, plus a copywriting suite, govern how agents plan, coordinate, remember, and write. Hot-reloaded, no restart.
Dynamic automation, everywhere
Put research, planning, scheduling, missions, email, and more on autopilot, any agent, any cadence. It ships with 24+ ready routines as a starting point: every one a plain, editable file you adjust, retarget, or clone into something entirely your own.
Documents & context artifacts
A documents workbench over your file corpus, plus curated context packages, the exact notes and files an agent needs, that pre-ground every new session by construction, and grow as your own project does.
Workbench comfort
A real desktop workbench: split-grid terminal workspaces, pop-out windows for every view and pane, three themes, a persistent activity log. Desktop quality, not a web shell.
Trust & control
Autonomy sells itself with big words. We sell the opposite: visibility, control, and the proof in the log, because that is what matters in production.
Drafts stay drafts
Anything outward-facing (a send, a publish) stays a draft until you release it. No surprise sends, ever.
Pause & audit instead of a black box
One click freezes every mutating agent action; every control action is a line in the audit log. Safety here is architecture, not a promise.
Local and private
Your files, memory, and knowledge graph never leave your machine. No usage analytics, no telemetry. The only built-in outbound call is a lightweight subscription check that sends no data and no files.
Built with the EU AI Act in view
The publishing and compliance layer is built with EU AI Act Article 50 in mind: automatic AI-disclosure on published pieces, plus an audit trail on every send.
Readable, editable doctrine
No black box: the agents, protocols, and configuration are plain, readable text on your disk. You can see how the system thinks, and you're free to change it for your own use.
Your subscription, your model access
The platform drives your own agent CLI: Claude Code or OpenAI Codex, your choice per dispatch. We never sit between you and your model access, and we never meter your usage.
Frequently asked questions
How is this different from an AI coding assistant?
Assistants wait for prompts. The Command Center runs missions: it decomposes goals into operations, staffs them with agent sessions in visible terminals, and reports when the success conditions are met, while you read along, intervene, and re-steer at any point. You stay the commander, not the clerk.
Where does it run, and where does my data live?
Entirely locally, on your machine. Files, memory databases, and the knowledge graph live on your disk; the services talk to each other, not to any cloud of ours. The one built-in outbound call is a lightweight subscription check that sends no usage data and no file contents.
What happens when an agent makes a mistake?
You see it. That is the point. Every session is a readable terminal, every control action is in the audit log, and one click pauses the entire fleet. And anything outward-facing, like a newsletter send, stays a draft until you release it.
Can I harvest my own sources?
Yes, that's the point. It ships with 100+ curated AI & tech sources so something runs on day one, and you can add unlimited sources of your own: any RSS/Atom feed, any public X account, any YouTube channel, or your own newsletter inbox, with no API keys. Paste a URL, test it live, schedule it.
Can it analyze YouTube videos?
Yes. Paste a URL or schedule a channel, and it produces an 8-section deep analysis with the charts, code, and links pulled off the screen and timestamped to the second. Batch several videos and it finds the cross-video trends.
What does Voice cost to run — and is dictation included?
Two answers, because two different things are being asked about. Dictation is included in Pro: an on-device speech model, fully offline, no account, nothing extra to pay. Voice — the conversational assistant that runs the platform hands-free — is Ultra. Its local side (opening views, arranging windows, driving terminals and walking your knowledge graph by speaking) costs nothing extra and works offline; its conversational AI runs on your own Grok subscription or an xAI/OpenAI key: your provider, your account, never metered by us.
Can I add my own agents?
Yes. An agent is a markdown file. Write one, or clone a shipped specialist in the app, and it's live on the next run. No restart, no rebuild. The same goes for skills, protocols, and scheduled routines.
What do I need to get started?
Windows 10/11 or macOS, the installer, and your own AI subscription (Claude Code or OpenAI Codex), plus Node.js and Python, which the installer sets up automatically if they are missing. The integrated graph database and all services are set up by the installer; we do not sell model access. Extra capabilities like semantic search and image generation use optional keys of your own.
What happens if I cancel?
Your data stays fully yours: files, memory, and the graph remain on your disk as open files you own. Only the execution of premium capabilities pauses; nothing is deleted, and nothing that belongs to you gets locked away.
AI Intelligence Newsletter
Curated AI insights, sent when there's something worth your inbox.