Let another agent work in your workspace

You ask Claude: "Which projects for Meyer GmbH still have unbilled hours, and who worked them?" Claude asks kyte, kyte answers from your company's data, and the reply names the projects, the hours and the people. Nothing was copied anywhere, nothing was pasted in, and you never left the program you were already working in.

That is what this does: kyte hands another AI program its data tools — to read, and to create, change and delete records. "Log three hours on the Meyer project for today" works the same way. The mechanism is MCP, a small protocol those programs use to pick up tools, and kyte speaks it as a command — kyte mcp.

What you need

The kyte command line. It is the same product as the app, without a window — see Install kyte for where it comes from. Sign in once:

kyte login

The browser opens, you confirm, the session is kept. After that kyte mcp needs nothing from you.

Connecting it

The other program starts kyte itself, so you tell it the command once. In a configuration file use an absolute path for a workspace folder: ~ is a convenience of the shell, and a JSON or YAML argument list is not a shell. One exception, and it is kyte's own: a leading ~ in --db is expanded by kyte itself since 0.1.1227, so that argument reads the same on macOS, Linux and Windows.

Claude in the terminal

claude mcp add kyte -- kyte mcp

Claude Desktop — install the kyte extension: download kyte.mcpb and open it, or drop it into Claude's Settings → Extensions. Claude shows what it is and asks you to install it; there is no file to edit. In the extension's settings you choose Read only and the data folder it works in — its own folder by default (~/.kyte/claude-desktop), so Claude Desktop and a kyte session in the terminal never compete for the one database. Sign in for that folder once:

kyte login --db /Users/you/.kyte/claude-desktop

The extension starts the kyte you installed; when kyte is updated, the extension uses the new one. Without kyte installed it still connects, and tells Claude — and so you — how to install it.

Hermes — one command, which writes ~/.hermes/config.yaml for you:

hermes mcp add kyte --command kyte --args mcp --db ~/.kyte/hermes

It gets its own data folder for the same reason Claude Desktop does: one folder serves one process, and a Hermes session and a kyte session in a terminal pointed at the same folder are two writers on one database. Sign in for that folder once:

kyte login --db ~/.kyte/hermes

Then give it the knowledge to go with the tools. Hermes hands a model tool names and descriptions and nothing else — what a server explains in its handshake is dropped on the way — so how to work in a kyte workspace is a skill of its own:

hermes skills install Pirol-ai/kyte-hermes/skills/kyte

The full address is the name here. Hermes answers hermes skills search from a central index that leaves GitHub out, so the short hermes skills install kyte finds nothing.

Grok — two commands, and they bring both halves at once: a Grok plugin carries the server and the skill that goes with it.

grok plugin marketplace add Pirol-ai/kyte-grok
grok plugin install kyte

The first adds our marketplace, the second installs the plugin from it and asks you to trust it — a plugin may bring hooks and servers, so Grok wants a clear yes. The data folder is its own, ~/.kyte/grok, and Grok expands that ~ itself before it starts the server. Sign in for the folder once:

kyte login --db ~/.kyte/grok

Then start a new session: the server is connected at session start, so a running one does not pick it up. grok mcp doctor says whether it came up, and grok inspect shows both halves — the skill and the server. grok mcp list does not: it reads config.toml alone and stays silent about anything a plugin brought.

To see what it offers before you connect anything, ask it by hand:

printf '%s\n' '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | kyte mcp

What it can read

The reading tools of the apps your workspace has switched on — seventeen when everything is on:

App It can list
CRM organisations, contacts, deals, products
ERP projects, organisations, time entries, and the hours booked on a project
Processes the process definitions
Goals objectives, key results, goals, initiatives, teams, people
Knowledge the company's documents
Tasks the board

It reads as you: the same rows your role sees in the app, no more. A colleague's private notes stay private, because they are private in the database and not merely hidden in the interface. An app you have switched off is not in a foreign agent's hands either.

What it can change

The same apps, the same records: it can create, change and delete them — a customer, a contact, a deal, a project, a time entry, an invoice, a goal, an absence, a document's entry. Two things stay reading only: the board's tasks, and the hours added up per project.

It writes as you, too. What your role may not change, it cannot change; a record it may not see, it cannot touch — and it is told so, instead of being told it worked. Every change is recorded as made by an agent on your behalf, in the same history the app shows.

Who asks you first is the other program: Claude asks before it uses a tool that changes something, and kyte marks every tool for what it does — reading, adding, or overwriting and removing — so that question comes where it belongs. Allow it once, or for the session, there.

One difference to kyte's own agent: a process that starts when a record changes starts for changes made in kyte, not yet for changes made over MCP.

Reading only

If an agent should read and nothing else, serve it that way:

kyte mcp --read-only

Claude in the terminal: claude mcp add kyte -- kyte mcp --read-only. Then the writing tools are not there at all — not hidden, absent.

The other agent is told what it has

When the connection opens, kyte introduces itself to the other program's model: whose workspace this is, that it works as that person under that person's permissions, that the data keeps itself current, what it may change and that every change is recorded as an agent's, and that apps are built in kyte itself. You do not have to explain any of that in your prompt.

It stays current by itself

With a team server the workspace synchronises while the server is serving, not only at the start. The other agent needs to know nothing about it and never has to ask for a sync: it asks a question, and the answer is on the company's current state.

If the box cannot be reached, kyte keeps answering from the local copy and says so once in its log. Yesterday's answer beats no answer.

What it can never do

  • Build. Apps and extensions are made in kyte itself — in the app or with the kyte agent, where the approval card is. That is a boundary, not a gap: see What Extend is.
  • Run commands or read your files. Only the data tools above exist for it.
  • Act past your permissions. It is you, as far as kyte is concerned — never more.

Where the data goes

This is the one thing that genuinely changes when you connect another program: what that program reads travels to that program's model, under that vendor's terms.

kyte's own agent sends through the Pirol endpoint, which replaces personal data before a message leaves the machine — see Personal data in a message. A foreign agent does not use that path. When Claude reads your customer list to answer a question, the customer list goes to Claude.

For many questions that is a reasonable trade. For some it is the wrong one. It is your decision — it should just be a knowing one.

One database, one program at a time

kyte's database takes a single connection. While kyte mcp serves an agent, do not have a kyte session open in another terminal — and two agents need two workspace folders, each with its own sign-in:

kyte login --db /Users/you/.kyte/claude
kyte mcp --db /Users/you/.kyte/claude

Each folder is its own workspace. Both show the same company only when both sync with a team server — see Alone, with Pirol Online, or on your own server. If a folder is busy, kyte says so and stops instead of opening a second connection.

And note that the command line's workspace is not the desktop app's: two folders on one machine. A team server is what makes them one company.

When something does not work

The other agent says kyte cannot answer right now — either the local session has run out (it lasts twelve hours) or the data folder is busy. The server stays connected in that case and offers a single tool that names the reason and the remedy, so ask the agent about it. Usually the remedy is kyte login, followed by reconnecting the server in the other program.

The agent still acts as the old user — it reads the session once, when it starts. After a kyte login, restart the other program or switch its kyte server off and on.

It sees old data — without a team server this folder stands alone, and its state is whatever is in it. With one, check the server's opening line: if it does not say it is syncing, the box cannot be reached.