Skip to content

MCP Server

3gpp-crawler mcp starts a stdio Model Context Protocol server (fastmcp) that exposes the crawler to AI agents.

Starting

3gpp-crawler mcp

Tools

Tools mirror the CLI groups:

Group Tool CLI equivalent
crawl crawl_tdocs tdoc crawl
crawl crawl_specs spec crawl
crawl crawl_meetings meetings crawl
crawl crawl_work_items work-items crawl
enrich enrich_work_items work-items enrich
query query_tdocs tdoc query
query query_work_items work-items query
query query_specs spec query
query query_meetings meetings query
checkout checkout_tdocs tdoc checkout
checkout checkout_specs spec checkout
checkout list_tdoc_files tdoc files
checkout list_spec_files spec files
main-file get_tdoc_main_file tdoc open (resolve only)
main-file main_file_set tdoc main-file set
main-file main_file_show tdoc main-file show
main-file main_file_list tdoc main-file list
main-file main_file_clear tdoc main-file clear
main-file reclassify_main_files tdoc main-file reclassify
tag tag_add tag add
tag tag_remove tag remove
tag tag_list tag list
database database_clear_tdocs database clear-tdocs
database database_clear_specs database clear-specs
database database_clear_meetings database clear-meetings
database database_clear_all database clear-all
database database_export_tags database export-tags
database database_import_tags database import-tags
config config_show config show

The database_clear_* tools delete records only. Downloaded TDoc and Spec files stay in the checkout directory unless purge_checkouts=true is passed; that is irreversible and rarely needed, because TDocs never change. | config | config_validate | config validate |

Notable parameters: query_tdocs accepts a tag list (comma = OR, repeated = AND, same as the CLI). crawl_tdocs, crawl_meetings, query_tdocs and query_meetings accept subgroup (e.g. ["SA4"], ["S4"], ["R1"]) alongside working_group. Working groups are only RAN/SA/CT — passing a subgroup code as working_group is rejected rather than silently collapsing to the plenary (SA4 → SA). crawl_tdocs accepts meeting (repeatable patterns: ["SA4#138"], ["S4-138"], ["SA4#13*"], or meeting IDs), mirroring tdoc crawl --meeting.

query_tdocs also accepts work_item (e.g. ["ULBC-MED"]) to select every TDoc whose meeting document list row references that 3GPP work item, and spec_number (e.g. ["26.114"]) to select TDocs that name that specification. crawl_work_items crawls the per-lead-body DynaReport work item lists; scope it with subgroup (e.g. ["S4"], also historical lead bodies like "GP") or working_group (an area: RAN/SA/CT), or omit both to discover every lead body. enrich_work_items fills status, parent/child and spec/CR links from the portal detail pages (default: TDoc-referenced and never-enriched work items; pass work_item/limit/include_all to steer). query_work_items reads the resulting catalogue: titles, releases, approval status, lead body, parent work item and impacted specifications, plus the number of TDocs referencing each work item (from the Related WIs column during crawl_tdocs). A database crawled before work item support holds none — rebuild it with database_clear_all and re-crawl.

The destructive database_clear_* tools refuse to run unless called with confirm=true — there is no interactive prompt over MCP, so the explicit flag is the confirmation (and costs agents one deliberate extra round-trip). Prefer database_export_tags first when a backup would help.

Main-document selection is available over MCP. get_tdoc_main_file downloads, extracts and resolves the main document but does not launch it — it returns the absolute path, the archive-relative member, and the source (manual_override, decision_model, heuristic). The main_file_* tools manage overrides, and reclassify_main_files replays classification for checked-out TDocs (tdoc_ids or all_tdocs=true); manual choices are protected unless include_manual=true, which deletes those overrides first. selector may be overridden per call. Selection settings live in the [checkout] section — configure them once in the global ~/.config/3gpp-crawler/config.toml so MCP and the CLI share them.

Client configuration

Add to any MCP-capable agent client:

{
  "mcpServers": {
    "3gpp-crawler": {
      "command": "3gpp-crawler",
      "args": ["mcp"]
    }
  }
}

When running from a checkout instead of an installed tool:

{
  "mcpServers": {
    "3gpp-crawler": {
      "command": "uv",
      "args": ["run", "--project", "/path/to/3gpp-crawler", "3gpp-crawler", "mcp"]
    }
  }
}

Output format

Tools currently return captured console text — fine for LLM consumption. For machine-parseable output, run the CLI directly with --output json.