Corpuskit

Config-driven knowledge corpus runtime + compression orchestration for AI coding agents — multi-agent via MCP, zero-config on any docs folder.

It turns a folder of markdown into a searchable, constraint-aware institutional memory your agent can query (instead of re-deriving from scratch), and orchestrates external token-compression tools. Generalized from a working Claude-Code system; a project is just one corpus.yaml.

What you get

• Manifest + FTS5 search over full document bodies (SQLite, stdlib — no embeddings/torch needed). Query with BM25.
• Machine-readable constraints — @constraint TYPE | target= | rule= | adr= | severity= lines in your decision/ADR docs, queryable by component.
• MCP server exposing kr_search / kr_constraints (names/descriptions configurable) — works with any MCP-capable agent (Claude Code, Cursor, …).
• Agent install — idempotent registration of the MCP server + write-back (SessionEnd) + constraint-injection (UserPromptSubmit) hooks.
• Compression orchestration — discover/install/lifecycle/health over external RTK (shell-output) + Headroom (transport proxy), used AS-IS (never vendored).

Install

pip install corpuskit            # core (CLI + manifest/index/constraints)
pip install "corpuskit[mcp]"     # + MCP server
pip install "corpuskit[all]"
# until published: pip install git+https://github.com/SupaKang/corpuskit

Quickstart (zero-config)

cd my-docs/                 # any folder of *.md
corpus index build         # project_key = top-level dirname; full-body FTS5 index
corpus index query "rate limiter design"
corpus constraints --component payments-api

No corpus.yaml needed — defaults to auto_layout + dirname keys + standard YAML/bulleted frontmatter.

Config (corpus.yaml) — opt in when you need it

knowledge:
  keyed_roots: { specs: spec, decisions: decision }   # relpath -> doc_type (key = subdir)
  flat_roots:  { daily: daily }
  frontmatter: { style: auto, fields: { project: [project], status: [status] } }
  constraints: { decisions_root: decisions }
agent: { type: claude-code }       # claude-code (MCP+hooks) | cursor (MCP) | standalone | cline/mcp-json
compression: { enabled: false }

corpus init scaffolds one. See examples/overmind.yaml for a full localized (Korean) instance.

Agent integration

corpus install --agent claude-code     # idempotent: MCP + SessionEnd + UserPromptSubmit (backs up settings.json)
corpus status --agent claude-code
corpus uninstall --agent claude-code

Restart your agent; it gains kr_search / kr_constraints tools, auto-injected active constraints, and self-updating index on session end.

Compression

corpus compression install   # ensure RTK + Headroom present
corpus compression start     # launch Headroom proxy, print ANTHROPIC_BASE_URL
corpus compression health    # versions, native-Windows degradation, port

CLI

corpus init | index build|query | constraints | serve-mcp | install|uninstall|status | compression … | doctor

License

Apache-2.0 (this code). External tools RTK and Headroom are separate Apache-2.0 projects, used as-is — see THIRD_PARTY.md.