noveletary

日本語

novel + secretary — a constraint-maintained narrative knowledge base and MCP server that checks the internal consistency of fiction, with first-class support for Japanese prose.

What

A local MCP server an LLM (Claude Code, Claude.ai Projects) calls while you write or import a novel. It tracks story facts in an append-only operation log, gates contradictions at write time, branches parallel plot drafts, and routes the questions it cannot decide to you — the author.

• Construction = checked mutation. Adding a fact runs hard-constraint checks (a dead character acting, a monotone ledger decreasing, a temporal cycle, an orphaning delete) and rejects contradictions with the conflicting fact set.

• Verification = the same engine, batch mode. Audit a whole branch; hard violations are certain, optional semantic checks (NLI) become author questions.

• Story branches are first-class: parallel drafts (A-plot / B-plot) audited independently, merged with structural conflict detection, rolled back without losing history.

• The author is the oracle. Unresolved aliases, merge conflicts, and semantic doubts go to the author, not to LLM guesswork. Answers persist and propagate through later checks.

Related MCP server: THOUGHT

Why

LLM-driven consistency checking is wasteful and unreliable when the LLM both writes and self-grades. noveletary keeps a deterministic constraint engine and the author as the trusted core, and treats the LLM as a fallible translator with no authority. Most "contradictions" in fiction are structurally decidable (state machines, numeric invariants, temporal constraints) and need no semantics; the semantic residual is the only part a language model judges, and even then the verdict is a question, not a gate.

Status

Early (v0.1). Core engine, store, tri-temporal facts (valid interval / discourse / transaction), branching, merge, audit, outline beats + foreshadow ledger, and the MCP server are implemented and tested. The Japanese NLP extraction layer (KWJA zero-anaphora + full noun-phrase reconstruction, GiNZA fallback) is the standard path but stays advisory — propose_canon_facts drafts canon from prose for author curation; it never gates. KWJA needs Python < 3.14 and self-seeds its checkpoint cache on first use. Not yet deployed remotely (Cloudflare Workers + D1 is a known migration path).

Install

pip install -e ".[dev]"          # core + tests
pip install -e ".[dev,nlp]"      # add Japanese NLP (GiNZA, KWJA)

Run as an MCP server

noveletary-mcp                                   # stdio
claude mcp add noveletary -- noveletary-mcp      # register with Claude Code

SQLite state persists in data/narrative.db (run from the repo root; override with NARRATIVE_DB).

Tools (LLM-facing)

Facts are tri-temporal: chapter is valid-time as an interval [chapter, valid_to) (when a fact is true in the story), narrated_in is discourse-time (which chapter reveals it — for foreshadowing / flashbacks), and the op-log is transaction-time. Each tool's description is prefixed with its category and carries read-only / destructive annotations.

To drive the tools from unattended agents / Claude Code subagents, allowlist the whole server (mcp__noveletary) rather than individual tools — see AGENTS.md.

License

MIT. See LICENSE.