Introduction
The MCP server gives a connected assistant a set of tools. You don’t call them yourself. You ask the assistant in plain language and it picks the right tool. This page lists what each tool does, so you know what an assistant can and can’t do with your silos. Every tool is limited to the silos owned by the account that created the MCP Key.Tools
A Read only key gets the five read tools. A Read & write key gets all nine.
Read Tools
list_silos
Lists your silos, newest first, 25 per page. Each entry has the silo’s UUID, name, size, visibility, region, and timestamps.
The result includes
has_more, which tells the assistant whether there is another page.
get_silo_meta
Returns a silo’s name, size in bytes, visibility, region, schema lock state, and timestamps. It does not return the data, so it is the cheapest way to check on a silo.
read_silo
Returns a silo’s full JSON data.
For silos over 100 KB, see Large silos.
query_silo
Reads part of a silo, using the same engine as Query and Search. The assistant gives exactly one of path, idx, or filters.
Each filter is an object with three fields:
For example, a filter that finds products with a price above 20:
get_silo_schema
Returns the JSON Schema attached to a silo, if any, and whether it is locked. An assistant can use this to learn what shape a write must have before it tries one.
An assistant can read a schema, but it can’t change, lock, unlock, or remove one. Schemas are managed in the console only. See Schema Lock.
Write Tools
These tools are only available to keys with Read & write access.create_silo
Creates a new silo from JSON data and returns its UUID. Your plan’s limits on silo count, silo size, and regions apply.
update_silo
Updates a silo’s name or visibility, replaces its data, or any combination of the three.
To change only part of the data,
patch_silo is the better choice, because the assistant doesn’t have to send the whole document again.
A silo’s region can’t be changed over MCP. Use the silo’s Settings tab in the console.
patch_silo
Changes part of a silo’s data with JSON Patch operations.
The supported operations are
add, remove, replace, move, copy, and test. Paths are JSON Pointers, such as /products/1/price. Use - to add to the end of an array, as in /products/-.
For example, these operations change one price and add a product:
- All or nothing. If any operation fails, none of them are saved.
- Checked like any other write. The patched data must fit your plan’s size limit and match a locked schema.
- Safe against conflicts. If the silo changes while the patch is being applied, nothing is saved and the assistant is asked to read the silo again and retry.
- Conditional changes. A
testoperation at the start makes the patch apply only if a value is still what the assistant expects.
delete_silo
Permanently deletes a silo and its data. This cannot be undone.
Resources
Silos are also available as MCP resources, for clients that let you attach a resource to a conversation.
The server lists up to 100 of your silos as resources, newest first. Reading a resource returns the same result as
read_silo.
Large Silos
Silos can be many megabytes, which is far more than an assistant can usefully read at once. When a result is over 100 KB, the server returns a summary instead of the data:
This applies to
read_silo, query_silo, and resource reads. With the summary in hand, the assistant can use query_silo with a path, idx, or filters to fetch exactly what it needs.
Errors
When a tool can’t do what was asked, it returns an error message that the assistant can read and act on, for example:Good to Know
- Reads are always fresh. MCP reads skip the cache, so an assistant that writes a silo and reads it back sees its own change.
- Writes are logged. Every create, update, patch, and delete over MCP appears in Logs.
- Data is not echoed back. After
update_silo, the assistant gets the silo’s metadata, not a copy of the data it just sent. - Silo data is treated as data. The server tells the assistant that silo content is user content and not instructions to follow.