Naski sits over Google Workspace, Microsoft 365, Slack, Salesforce and thirty other tools, and runs agents on a schedule or on an event — with a human approval gate wherever it matters.
Ask for an answer, or create an agent that keeps working on it
Naski does not ask a company to move its work. It reaches into the tools the work already lives in, and puts a reviewable agent in front of them.
Install a connector from a reviewed catalogue. The credential is stored encrypted and injected server-side, so agent code can name a call but never perform an arbitrary one — and never sees the key.
An agent fires on a cron schedule, at a fixed interval, on an inbound email, on a posted webhook, on a file added to the shared space, or on a form someone sent.
An agent can pause until a person approves — before it writes to a customer, spends money or changes a record elsewhere. No answer means no: the safe branch is always the one that does nothing.
An agent can belong to the organization rather than to whoever wrote it. Its author can change role or leave; the agent keeps running, on the company's credits and within the budget set for it.
A reviewed catalogue that ships with the product and installs in one gesture from Settings — not an integration built per customer.
13 of them connect by pressing Connect and approving on the provider's own screen, where that provider is enabled on your deployment. The others take a key you paste in Settings.
See the full catalogueThree steps, and nothing hidden between them.
Press Connect and approve on the provider's own screen, or paste a key in Settings. Keys live in Settings, encrypted — never in a chat message.
Tell the chat what should happen and when, in plain language. It writes the agent; you read what it does, and what it may touch, before saving.
Every run leaves a result, a journal in plain prose and a record of each decision. Grants, credential changes and document reads land in an audit trail you can export.
Written against what the engine does today, not against a roadmap.
Supplier documents arrive by email or land in the shared space. The agent reads them — PDF, Word, Excel, or a scan where OCR is enabled — extracts the fields you named, and upserts one row per reference. Anything failing a check goes to a review queue instead of going through, and a person decides on it later.
One cron line and a time zone. Every Monday at eight the agent pulls what it needs from Stripe, Qonto or Pennylane, has a model turn it into a short brief, and posts it to your Slack or Teams channel. It keeps a journal of what it already reported, so the next brief is about what changed.
Publish a page with a form. Sending it stores the attachment, wakes an agent that reads the submission, writes the result back onto that same page and notifies your channel. Only the owner can submit — a shared link stays read-only.
The chat writes the agent. You read it before it runs.
Try a prompt
Create an agent that files every supplier invoice from my inbox
Naski keeps the interface simple while each job gets the time, the context and the permissions it needs.
No code, just say it
Bring your own context
Keep work moving
Ownership, budgets and evidence — decided before an agent is allowed to act, not reconstructed afterwards.
Cap what one agent may spend over a rolling window, and cap how much budget any one person may hand out. Both are checked before a run and before each billed step.
Every access to a shared agent goes through an explicit view / run / edit check. A company agent with no team is reachable by its administrator and the organization's admins, and by nobody else.
Grants, role changes, credential settings, approval decisions and document reads are recorded with who, what and when — and downloaded as CSV.
Start on chat, then add agents, connected tools and external AI clients as more of the work moves in.
Try the workspace
Use credits for model-assisted work. ≈5 agent builds.
Agents available in your workspace.
Bring context and actions from connected tools.
Ask and act directly from chat.
Let work run again without another prompt.
Keep outputs and run history available.
Connect compatible AI clients through MCP.
For regular workflows
Use credits for model-assisted work. ≈25 agent builds.
Agents available in your workspace.
Bring context and actions from connected tools.
Ask and act directly from chat.
Let work run again without another prompt.
Keep outputs and run history available.
Connect compatible AI clients through MCP.
For growing teams
Use credits for model-assisted work. ≈50 agent builds.
Agents available in your workspace.
Bring context and actions from connected tools.
Ask and act directly from chat.
Let work run again without another prompt.
Keep outputs and run history available.
Connect compatible AI clients through MCP.
For demanding work
Use credits for model-assisted work. ≈100 agent builds.
Agents available in your workspace.
Bring context and actions from connected tools.
Ask and act directly from chat.
Let work run again without another prompt.
Keep outputs and run history available.
Connect compatible AI clients through MCP.
What a company usually asks before letting software near its tools.
A workspace where a company's people chat with an AI, and where the workflows worth repeating become agents that run on a schedule or on an event, against the tools the company already uses.
Tell us which tools your team lives in and what the recurring work looks like. A person answers, at a real address.