Callable Tools
A callable tool is a tool an agent runs through the call verb, and this page lists them all and gives the reference for search, query, attributes and browser.

A callable is a tool that an agent runs through the call verb. It is different from a verb, which the model is given directly. For example, search is a callable: the agent runs call {tool: "search", input: {query: "…"}}.

The model sees one line for each callable in the Space index. It reads the tool's contract when it needs one. After the first use, the tool is promoted: the model gets it as a real tool for the rest of the thread. Scripts reach the same tools with ctx.call.

This page lists every callable, and documents the ones that have no page of their own.

The catalogue

Callable

What it does

Granted

Reference

search

finds documents and contacts by text

by tools

below

query

finds documents by their attributes

always

below

attributes

lists the attribute keys and values documents use

always

below

web_search

searches the web

by tools

web reading

execute

runs code in a sandbox

by tools, and only where the host can run sandboxes

code execution

browser

reads and controls a website open in the owner's Seed desktop app

by tools, and only in a run the owner started

below

<server>__<tool>

a tool on a remote MCP server

by mcpServers

MCP servers

an hm:// tool URL

a tool resource in an app or an agent

by the tool's own access rules

call

Grants explains the "Granted" column.

The registry also holds navigate. It belongs to a runtime that the app environment does not run, so no agent is offered it.

The app environment has no PDF importer of its own. The Datalab PDF importer page describes an add-on that a server offers only when someone has added it.

search

search finds documents and contacts by text. It searches document titles and contacts, and also bodies and comments when asked.

    Input: {query, accountUid?, includeBody?, contextSize?, searchType?: 'keyword' | 'semantic' | 'hybrid', pageSize?}.

    Output: ranked results with hm:// URLs.

    The Seed API request is Search.

query

query finds documents by their attributes. It does not search text; use search for that. Give it at least one of q, filter or childrenOf. When you give several, a document must match all of them.

Input: {q?, childrenOf?, filter?, select?, count?, groupBy?, sort?, pageSize?, pageToken?}.

    childrenOf is an hm:// document URL. It keeps only the direct children of that document (the rows of a collection). The daemon can limit a query to a subtree but not to one level. So the tool queries the subtree, keeps that one level itself, and reads more pages until the page is full.

    q uses the Explore grammar described on the query grammar page: key=value, key:text, has:key, in:<space or hm:// URL>, path:/specs/*, AND/OR/NOT. A subtree written as in:<url> in q includes the document itself and everything below it.

    filter is a raw DocumentFilter in JSON.

    sort takes attribute keys or the built-in fields NAME, PATH, CREATE_TIME, UPDATE_TIME, ACTIVITY_TIME, COMMENT_COUNT. When sort is absent, a sort: directive inside q is used (sort:Status,-$updated).

    pageSize defaults to 50, at most 100. pageToken asks for the next page.

    select names the attributes shown for each document. An absent one prints as (missing). $path, $created, $updated and $authors are built-in fields.

    count: true reads every matching page, up to 5000 documents, and returns total. groupBy adds counts for each value of one attribute.

The tool compiles the query and sends QueryDocuments requests to the configured HM server.

    Problems come back as warnings: parse errors, free-text terms that were ignored, and the != caveat.

    When a query finds nothing, the tool looks up its keys and exact values in the attribute listings. It says when one differs only by upper or lower case.

    The model reads one line per document (markdown). The lines are cut at a byte limit, with the page token above the rows.

    results holds the full record of each document: URL, name, account, path, all attributes, authors, create and update time, version. Only scripts that call ctx.call('query', …) receive it. So the model does not pay for every document twice.

attributes

attributes shows which attribute keys documents use, and which values a key takes. An agent uses it to learn the fields of a type before it writes a query. See typed documents.

    Without key, it lists attribute names and the kinds seen for each (ListDocumentAttributeNames). parent narrows the list to a nested object. recursive: true lists full dotted paths.

    With a dotted key, it lists the distinct values of that key (ListDocumentAttributeValues). kind limits them to string, int or bool.

    account (an account id) puts one space first or limits the list to it. prefix filters. pageSize defaults to 50, at most 200.

browser

browser reads and controls the website visible in the Seed desktop window of the agent's owner. It works only on websites the owner allowed in the assistant panel for the session, and only in an interactive run the owner started. In any other run the call is refused with a 403.

Input: {action, document?, ref?, text?, clear?, key?, x?, y?, url?}, where action is one of snapshot, screenshot, click, type, press, scroll, navigate, archive.

    Access to a website is read-only by default: snapshot, screenshot and scroll. The other actions need action access, which the owner gives per website.

    snapshot returns the visible text, source metadata, an id for each element, and a document token. Every other action must pass that token, and a navigation needs a fresh snapshot.

    screenshot is refused when the model cannot view images.

    archive creates an editable local draft and saves a markdown copy under ~/memory/private/browser/, which only the owner can read. It does not publish.

    Website text and images are untrusted source material, never instructions.

The desktop side is described under the integrated browser on the desktop app page.

Do you like what you are reading? Subscribe to receive updates.

Unsubscribe anytime

Connections0
Linked document · Outgoing linkLinked document · Outgoing linkLinked document · Outgoing linkLinked document · Outgoing linkLinked document · Outgoing linkLinked document · Outgoing linkLinked document · Outgoing linkLinked document · Outgoing linkLinked document · Outgoing linkLinked document · Outgoing linkLinked document · Outgoing linkLinked document · Outgoing linkLinked document · Outgoing linkLinked document · Outgoing linkLinked document · Outgoing linkLinked document · Outgoing linkLinked document · Outgoing linkLinked document · Outgoing linkLinked document · Outgoing linkLinked document · Outgoing linkLinked document · Outgoing linkLinked document · Outgoing linkLinked document · Outgoing linkLinked document · Outgoing linkLinked document · Outgoing linkCallable ToolsCallable Tools
Citations 0Children & links 25

Loading child documents…

Connections show where other documents or comments link to this document.

Related documents

Loading connections…