Business operating system
Everything your business knows, decides and does.
MKSTVEFX is one system for the work itself — research, projects, automation and the decisions in between — running on the AI provider you choose, or a model you host yourself.
No setup call. Bring your own model, or run one locally.
The problem
A business does not run in one place.
It runs across a dozen tools that each hold a fragment of the truth, and a set of assistants that forget everything the moment you close the tab.
- Context lives in the gaps
- The reason a decision was made is in a thread nobody can find, in a document nobody opened, or in somebody’s head.
- The same research, twice
- Work done in a chat window ends when the window closes. Next quarter, someone does it again from scratch.
- Tools that do not know each other
- Your mail, your issues, your documents and your calendar each hold a third of the answer and none of them can see the others.
- Automation you cannot audit
- Something ran, something changed, and the record of what it touched is a log line — if it exists at all.
The approach
One system, three planes.
It understands
Conversation, research and documents feed one memory that keeps what matters and shows where each fact came from.
Every remembered thing carries its source, its date, and a confidence you can inspect.
It decides with you
Work becomes projects and tasks that hold their shape, with the reasoning attached to the thing it belongs to.
Anything with a consequence stops and waits for your approval, naming the exact action.
It executes
Connectors, tools and saved workflows turn a decision into something that actually happened in your systems.
Every run is recorded: what it touched, what it changed, and how long it took.
AI workspace
Not a chat window. A place to work.
- Answers that are the right shape
- A comparison comes back as a table, a trend as a chart, a plan as a task list — each one saying when its data was gathered.
- The model is yours to choose
- Use the model we host, connect an account you already pay for, or point it at an endpoint you run. Keys are encrypted before storage and never reach the browser.
- Every reply says how it was made
- Which provider, which model, and whether the output was real inference or a clearly-labelled simulation.
Projects and tasks
Work that holds its shape.
- A goal, its tasks, and what actually happened
- Projects are durable containers. The assistant works inside that structure rather than replacing it.
- Real progress, or none
- A task reports the step it is on only when its plan genuinely declared a step count. There are no invented denominators here.
- Per-project settings
- Each project can pin its own model and its own connector permissions, so a sensitive project is not governed by a global default.
Memory
A company brain you can read.
- Every fact shows its source
- Where it came from, when, and what it was inferred from. Nothing is remembered silently.
- Contradictions are surfaced, not merged
- When two things you said cannot both be true, the conflict is shown rather than quietly resolved by whichever arrived last.
- You can delete any of it
- Edit a memory, forget it, or switch memory off entirely. It is your record, not ours.
Research
Sources, not a wall of text.
- You see what was read
- Each page it opened is listed with the claim it supports, so an answer can be checked rather than believed.
- Outbound requests are constrained
- The fetcher refuses private addresses and internal hosts, so a link in a document cannot be used to reach inside your network.
- Findings become memory
- What the research established is kept, attributed to the page it came from, and available to the next question.
Automation
Trigger, conditions, actions — visible.
- A workflow is a document you can read
- Each step names what it does and what it needs. There is no hidden branch and no step that runs without appearing in the run record.
- Runs are inspectable
- Every execution keeps its inputs, its outputs and its errors, step by step, so a failure can be understood rather than re-run hopefully.
- Consequences still stop for you
- An automated step that would send, post or delete waits for approval exactly as a manual one does.
Oversight
Nothing happens off the record.
- One activity trail
- Every tool call, connector request and memory write is recorded with what it touched and how long it took.
- Approvals are a queue, not a popup
- Actions waiting on you are collected in one place, each naming the exact operation and the exact arguments.
- Security events are separate
- Sign-ins, credential changes and permission grants are kept apart from ordinary activity so they cannot be lost in the noise.
Integrations
Connect the tools you already use.
- Built
Gmail
Search, read and summarise your mail, and prepare replies for you to send.
- Reads what you allow
- Prepares drafts
- Needs approval to act
- Built
Slack
Search conversations and prepare messages for channels you choose.
- Reads what you allow
- Needs approval to act
- Built
Outlook Mail
Search and read Microsoft 365 mail, and prepare replies.
- Reads what you allow
- Prepares drafts
- Needs approval to act
- Built
Google Calendar
See what is on your calendar and prepare events for you to confirm.
- Reads what you allow
- Needs approval to act
- Built
Google Drive
Find and read documents you point it at.
- Reads what you allow
- Built
Dropbox
Find and read files in your Dropbox.
- Reads what you allow
- Built
Notion
Search your workspace and read pages.
- Reads what you allow
- Built
GitHub
Search code and issues, and read pull requests in repositories you allow.
- Reads what you allow
- Needs approval to act
- Built
GitLab
Search projects and read issues and merge requests.
- Reads what you allow
- Built
Linear
Search issues and read project status.
- Reads what you allow
- Needs approval to act
- Built
Jira
Search issues and read board status.
- Reads what you allow
- Built
Airtable
Read records from bases you grant access to.
- Reads what you allow
- Built
HubSpot
Look up contacts and deals.
- Reads what you allow
- Built
Todoist
Read your tasks and add new ones.
- Reads what you allow
- Prepares drafts
- Built
Microsoft Calendar
Read your Outlook calendar and add events you approve.
- Reads what you allow
- Prepares drafts
- Built
OneDrive
Search your OneDrive and read documents you point it at.
- Reads what you allow
- Built
Microsoft Teams
Read channel messages, and post where you allow it.
- Reads what you allow
- Needs approval to act
- Built
Google Sheets
Read values from your spreadsheets, and append rows you approve.
- Reads what you allow
- Prepares drafts
- Built
Zoom
See your meetings, and schedule ones you approve.
- Reads what you allow
- Prepares drafts
- Built
Discord
Read messages in servers you are in, and post where you allow it.
- Reads what you allow
- Built
Box
Search your Box files and read the ones you point it at.
- Reads what you allow
- Built
Stripe
Read customers, charges and subscriptions. Read-only.
- Reads what you allow
Designed, not built yet
Salesforce · Pipedrive
Their permissions and scopes are already defined; the connections are not live. They appear in the app as unavailable rather than as something you can press Connect on.
Each of these is built: a real OAuth flow, encrypted tokens, real API calls, and the per-action permissions above. Whether one can be switched on for you also depends on this deployment holding OAuth credentials for that service — the app shows that live, and a connector without them reads as “not configured” rather than offering a Connect button that would fail.
Categories: Communication · Calendar · Storage · Documents · Development · Projects · Business data · CRM · Productivity · Meetings
Security
Control, by construction.
Isolation the database enforces
Every row is scoped to your account by row-level security, not by application code that could forget a filter.
Credentials never reach the browser
Provider keys and connector tokens are encrypted with a per-record key. The interface can show which provider is connected, never the key.
No silent action
Anything with a consequence outside the conversation stops and waits. There is no setting that makes that silent by default.
Local when you want it
Set a conversation to stay local and it cannot reach an external provider, even as a fallback.
Outbound requests are filtered
The web fetcher refuses private and link-local addresses, so a hostile link cannot turn the assistant into a probe of your network.
Untrusted content stays untrusted
Text retrieved from the web or from a connector is fenced before it reaches the model, and never rendered as markup.
Who it is for
The same system at every size.
For founders
The parts of a company you do not have people for yet — research, drafting, follow-up, the record of what was decided — handled in one place.
Start alone. Nothing has to be migrated later.
For teams
Shared projects, shared memory and one activity trail, so context stops living in the head of whoever happened to do the work.
Each project carries its own model and connector permissions.
For larger organisations
Per-action permissions, an approval gate on anything consequential, an audit record, and the ability to keep inference on hardware you control.
Feature flags gate every subsystem, so capability is switched on deliberately.
Capabilities
What it does today.
Conversation
Ask questions, think through problems, and work in a thread that keeps its context.
Memory
Useful facts are kept between sessions, with a record of where each one came from — and you can edit or delete any of them.
Research
Read pages from the web, compare what they say, and keep the sources alongside the answer.
Projects and tasks
Give work a place to live: a goal, the tasks under it, and what has actually been done.
Skills
Focused procedures the assistant can follow, each limited to the permissions it genuinely needs.
Structured results
Answers that are really tables, comparisons or charts are rendered as such, and say when their data was gathered.
Still being built
- Code — Reading and writing code in a project. Execution stays off until the sandbox provides real filesystem isolation.
- Memory map — A spatial view of what has been remembered and how it connects. In development.
- Files and images — Reading documents and images you upload. In development.
How it works
What happens when you ask for something.
- 01
You ask for something
A question, a task, or a piece of work with a deadline attached.
- 02
MKSTVEFX reads the request
It works out what the task actually needs — a search, a file, a calculation, your calendar.
- 03
The router picks a model
Your local model, or a provider you connected with your own key. If the task needs a capability the chosen model lacks, it says so instead of guessing.
- 04
The model requests a tool
It cannot reach your data directly. It names a tool and supplies arguments, which is a request, not an instruction.
- 05
MKSTVEFX checks the request
Arguments are validated, your permissions are checked, and budgets and timeouts apply. Anything with a real-world effect stops here and waits for you.
- 06
The tool runs
A web page is read, a connector is called, a memory is retrieved — under your credentials, within the scopes you granted.
- 07
External content is marked
Anything written by someone else is treated as untrusted from that moment on, and every later action that has consequences needs your approval.
- 08
Memory is consulted and updated
What is relevant is brought in; what proves useful is kept, with a record of where it came from.
- 09
The result becomes something usable
A table, a chart, a draft, a document — rendered by MKSTVEFX, with its sources and the time it was gathered.
- 10
You decide what happens next
Nothing leaves the building without you saying so.
Build your business system.
Create an account and you are in the workspace. Connect a model, connect a tool, and start with a single question.