Limits and errors

The limits that keep tool calls fast and predictable, and what an AI client sees when a call fails.

Limits per call

What	Limit
Command output (`exec`, `shell_run`)	32,000 bytes; longer output is cut
Command timeout	60 seconds by default, 900 seconds at most; use `job_start` for longer work
Machines per `exec_many`	8, each with half the output limit
`read_file`	200,000 bytes by default per call, up to 4,000,000; continue from `next_offset`
`write_file`, `copy_between_hosts`	4,000,000 bytes; use `transfer_start` for larger files
`transfer_start`	No size limit; 2 at a time per connection, up to 10,000 files each
Job output (`job_output`)	64,000 bytes per read
Terminal output per call	20,000 characters; waits up to 600 seconds
Port-forward HTTP body (`tunnel_http`)	32 KB

Open sessions

What	Limit
Shell sessions	4 per machine; closed after 30 minutes idle
Terminals	8 per connection; closed after 30 minutes idle
Tunnels	Closed after 1 hour
Schedules	200 per account
Web shares	24 hours by default when an AI creates one, 30 days at most
Waiting approvals	20 per connection

Calls that take longer than 10 seconds send MCP progress notifications every 10 seconds when the client asks for them, so the client does not time out while a command, a transfer or an approval is pending.

Plan limits

Plans set how many machines and members an account has, and which features its policies can use, such as per-folder file rules, command rules, schedules, web shares and secret environment variables. See pricing. A call that needs a feature the plan does not include fails with a message saying so.

Errors

A refused or failed call returns an MCP tool result with isError: true and a plain-language message that tells the AI what happened and what to do next. The messages you are most likely to see:

Message begins with	Meaning
`token lacks fleet.…`	The connection was approved without the scope this tool needs. Reconnect and approve the scope.
`refused by this connection's command rules`	The policy's command rules deny this command.
`high-risk operation`	The AI must show you the exact command, then call again with `confirm=true`.
`refused by guardrail`	A command LetKite never runs, such as wiping a disk.
`waiting for a person to approve this`	An approval is pending; the message says who was asked and until when.
`this connection's authorization is no longer valid`	The connection was disconnected. Reconnect the client to sign in again.

Sign-in errors use the standard OAuth error codes (invalid_grant, invalid_client, access_denied and so on). A missing or expired token gets HTTP 401 with a WWW-Authenticate header, and clients sign in again or refresh by themselves.