Skip to content

Practical use cases

These examples use the current MCP surface. They are patterns an agent can follow, not server-side automations.

Maintain a runbook during an incident

  1. Find the affected service with search.
  2. Read the relevant section through concept_read.
  3. Patch the verified recovery step with concept_patch and its if_match hash.
  4. Add a timestamped incident note with log_append(path=...).
  5. Run lint on the runbook scope and review the git commit.

Cartographer does not infer the root cause or open a review PR. The agent and operator own those decisions.

Keep operational procedures current

Use changes_since to see what changed recently, then update a procedure with concept_patch or replace it with concept_write. supersede can mark one concept as replaced by another; git retains the full history. if_match prevents an older session from overwriting a newer revision.

Import an existing notes directory

cartographer import creates resumable status: imported drafts. Work through them incrementally:

  1. inventory with concept_list;
  2. search/read a bounded group;
  3. curate types, titles, links and provenance;
  4. clear the imported status when a concept is reviewed;
  5. run validate and lint.

The bundled kb-import skill contains the operator procedure.

Share one KB with different agent clients

Run the HTTP server, then use cartographer connect for Claude Code, Codex, Kiro or OpenCode. The client installs one MCP entry per KB and materializes the KB's managed skills, agents, hooks and instructions. cartographer status shows drift; cartographer sync reconciles it.

Providers differ in what they can represent. See configurator and synchronization for the exact mappings and limitations.

Separate domains into multiple KBs

Mount independent git repositories under one HTTP server and give tokens kb:<name>:r or kb:<name>:rw scopes. A client connects to a specific KB through its query/path route, and each KB has independent indexes, commits, sync state and conflict registry.

Authorization is per KB, not per concept. Use separate KBs when data needs different access boundaries.