Documentation

How to actually use it

The three things a new team needs on day one. Everything here describes the product as it ships today.

01

What an agent is, and how to ask for one

An agent is a piece of TypeScript written against a single object. You do not have to write it: describe the job in chat and Naski writes it, then shows you the code before anything is saved.

  1. 1

    Say what should happen, and when. “Every weekday at 8am”, “whenever a PDF lands in a folder”, “when someone emails this address”. The schedule and the trigger are read out of the code without running it, so they are written as fixed values.

  2. 2

    Read what it does before saving. The code is the behaviour — there is nothing hidden behind it.

  3. 3

    Give it a folder if it should only see part of your space. Its file tools are then confined to that subtree.

  4. 4

    Let it keep a journal. An agent writes one plain-language line per thing it did, tagged pending or done, and reads them back on the next run. That is how a long job survives being split across many short runs.

  5. 5

    Decide what needs you. Asking for approval pauses the run until a person answers and treats silence as a refusal; filing a review queues an item for later without blocking.

  6. 6

    Editing an agent creates a new one, so the version that worked stays recoverable.

Worth knowingAgents have no ambient network access. Everything they reach — the web, your space, a connected tool — goes through a named call the platform performs on their behalf.

02

How to connect a tool

A connector is data, not code: a base address, an auth shape and a list of operations. Naski builds the request and injects your credential server-side.

  1. 1

    Open Settings → Connectors and install from the catalogue. You can also ask the chat to install one for you.

  2. 2

    Where the provider is registered on your deployment, press Connect and approve on the provider's own screen. No key is copied anywhere.

  3. 3

    Otherwise, paste the key in Settings. Never paste a key into a chat message: keys belong in Settings, where they are stored encrypted and injected server-side.

  4. 4

    Some services live at your company's own address — Salesforce, Zendesk, Jira, Confluence, Zoho CRM, Odoo, Shopify, DocuSign. Installing those asks for your instance URL and refuses without it.

  5. 5

    Outbound calls go over HTTPS only, and are refused when the address resolves to a private or reserved network.

  6. 6

    If a provider refuses to refresh a token, the connection is cleared and you are told to reconnect — rather than every call failing quietly.

Worth knowingA service outside the catalogue can still be reached: the chat can author a connector for it in the same declarative shape, and a remote MCP server covers tool-shaped services.

03

How to connect an external AI client

Naski exposes its own MCP server, so a client such as Claude Code, Claude Desktop or ChatGPT can list, create, run and inspect your agents from where you already work.

  1. 1

    Open Settings → MCP servers. The endpoint is shown there, and it is this one:

    https://api.blackbox.finance/mcp
  2. 2

    Claude Desktop and ChatGPT connect over OAuth: add the endpoint as a connector and approve. There is no token to copy.

  3. 3

    For a command-line client, generate a token in Settings and run:

    claude mcp add --transport http naski https://api.blackbox.finance/mcp --header "Authorization: Bearer YOUR_TOKEN"
  4. 4

    A token is shown once and stored only as a hash. Revoke it from the same screen; an account may hold up to twenty.

  5. 5

    External AI connections are part of the paid plans — the free tier does not include them.

Worth knowingEverything an MCP client sees is scoped to the account whose token it uses: that account's agents, runs and credits, and nothing else.

Ready to try it on a real workflow?

The free tier includes one active agent and one scheduled run — enough to see whether the shape fits before anyone signs anything.