Kraterion

Agents

Tools

Tools are the actions an agent can take during a turn. You choose which ones it's allowed to use; the model decides when to call them.

Enabling tools

List the tool names in the agent's tools field. A tool is only useful if the agent also has what it needs: kraterion_search requires a knowledge-enabled bucket, and the read/write tools require the relevant bucket to be attached and granted to the agent's sub-wallet.

The eight tools

ToolKindArgsDoes
kraterion_searchreadbucket, query, top_k?Hybrid keyword + vector search over a knowledge-enabled bucket.
kraterion_list_bucketsreadList the buckets attached to the agent.
kraterion_list_objectsreadbucket, prefix?, limit?List object keys in a bucket.
kraterion_read_objectreadbucket, keyRead an object's contents.
kraterion_write_objectwritebucket, key, content, content_type?Create or overwrite an object.
kraterion_get_manifestreadbucket, keyFetch an object's knowledge manifest (chunks, hashes, blob ids).
memory_rememberwritecontentSave a fact to the agent's memory.
memory_recallreadqueryRetrieve relevant facts from the agent's memory.

Read vs write

Read tools observe your data and have no side effects. Write tools change it: kraterion_write_object stores a new object — which encrypts the bytes, uploads them to Walrus, and records the result on-chain — and memory_remember appends to the agent's memory. Grant write tools only to agents you intend to let modify storage.

Tool-call trace

Every tool the agent calls during a turn is surfaced in the chat response, under the kraterion.tool_calls field (and as streaming kraterion.tool_call frames). Write tools that touch storage include the resulting on-chain transaction and Walrus blob ids, so an answer that changed something is traceable end to end. See the Chat API for the exact shape.