Document navigation,
for AI Agents.

# Installed as an MCP server, docslicer exposes five tools.
# Your agent calls them directly — no Python, no glue code.

parse(source, password=None, refresh=False)
# Local path or http(s) URL — format detected automatically.
# → doc_id, title, pages, document_tokens, and the outline:
#
#   - Highlights                        ~2k
#   - Financial Summary                 ~48k
#     - Income statement commentary     ~12k
#       - Net interest income           ~3k
#
# Every line carries what reading it would cost, subsections included,
# so the agent budgets context before spending it. Short documents skip
# the outline and come back as full text.

search(doc_id, query, limit=8)
# For when the outline is not enough — a heading that names nothing
# ("Note 14"), or a figure buried in a table no heading mentions.
# → hits: heading, snippet, read_tokens. Places to read, not answers.

read(doc_id, headings)
# → text, tokens_returned
# Headings exactly as the outline prints them; a heading carries its
# subsections. "Notes > Revenue" picks one when the name repeats.
# [Page X] markers stay inline, numbered as the document numbers them.

get_outline(doc_id)
# The outline again, once it has scrolled out of context. Cheap.

to_markdown(source | doc_id, output_path=None)
# Writes the whole document to disk and returns output_path, never the
# text — so document size never touches the agent's context.