Skip to content

Talamus Agent Tool Calling Guide

Agents should prefer the MCP server when available: it exposes read tools for grounded context and write tools that preserve review/provenance rules.

MCP server

talamus mcp install
talamus mcp serve --root .           # stdio
talamus-mcp --http --host 127.0.0.1 --port 8000

MCP tools

Tool Purpose
search(query, smart=False) Find relevant notes by title and summary; smart=True adds cached LLM query expansion.
read_note(title, as_of="") Read the full Markdown note, optionally as it existed at a past time.
ask(question) Get a cited answer from the configured brain.
verify(title) Check one note against its preserved source and propose a correction if needed.
recall(question) Return raw relevant note context for the agent to reason over.
overview() Show the domain map for orientation.
neighbors(concept) Show typed graph/ontology neighbors for a concept.
history(title) List past versions of a note.
sources(title) Show recorded provenance for a note.
ontology_status() Report schema version, type counts, and typed-edge coverage.
remember(text, scope="project") Save an important session insight into the project or central brain.
ingest_text(text, name="insight", scope="project") Compile selected text into notes without the worth-remembering gate.
propose_note(text, reason="") Put uncertain knowledge into the review queue instead of writing it directly.
review_list() List pending review decisions.
review_apply(item_id) Apply a review item while preserving history.
review_reject(item_id, reason="") Reject a review item and keep the decision logged.

CLI equivalents

When MCP is unavailable, wrap these commands as tools:

Command Purpose
talamus search "<query>" [--smart] Find candidate notes.
talamus read "<title>" [--as-of T] Read real note content before answering.
talamus recall "<question>" Retrieve context without spending an answer call.
talamus ask "<question>" [--as-of T] [--trace] Get a cited answer and optionally inspect the route.
talamus verify "<title>" Check one note against source provenance.

The graph and search indexes are routing aids, not answer sources. Agents should answer from read_note, recall, or ask output and keep citations.