Scrivener MCP Server

Scrivener MCP

MCP server for Scrivener: connect your novels, screenplays, and manuscripts to Claude, ChatGPT, and any AI assistant, with document management, writing analysis, content enhancement, offline semantic search, and character/plot tracking.

Start with the safe first session, then use editorial review and workflows, export presets, Content Credentials, or the Windows guide. Compatibility evidence records actual coverage and limits.

---

Scrivener MCP lets your AI assistant open, read, edit, analyze, and search your Scrivener projects directly. No exporting, no copy-pasting text back and forth. Tell it which project to open and start working.

You: Open my novel and analyze the pacing in Chapter 12.

Claude: Opens the .scriv project, reads Chapter 12, runs pacing analysis. The first half moves well, but the middle stalls — a three-page internal monologue starting at paragraph 14 kills the momentum you built in the confrontation scene. Consider cutting it to a paragraph and moving the backstory to Chapter 8.

Works with Claude Desktop, Claude Code, VS Code (Copilot/Continue), Cursor, and any MCP-compatible client. Reads Scrivener 3 projects on the supported platforms. Listed on the official MCP Registry as io.github.writerslogic/scrivener-mcp.

Install

Mac app: Draftbridge for Scrivener, $4.99

Not a terminal person? Draftbridge for Scrivener is the ready-made Mac app. It finds your AI apps — Claude Desktop, BoltAI, Cursor, VS Code, LM Studio, Raycast, Claude Code — connects Scrivener MCP to each with one click, shows your projects, and repairs the connection after an app update. No Terminal, no Node.js, no settings files. One-time $4.99, up to 3 Macs, signed and notarized by Apple.

The server, setup tool, and Claude Desktop extension here stay free and open source. The app is closed source; the fee buys the notarized build and supports development. See the Mac install guide and FAQ. Report app problems with the mac-app label.

Free: Claude Desktop extension or command line

Prefer to do it yourself? The .mcpb extension (Claude Desktop only) and the terminal setup below do the same job for free.

From a terminal:

npm install -g scrivener-mcp

Restart Claude Desktop and it's ready. Other clients need one more step:

npx scrivener-setup

This finds Claude Desktop, Claude Code, Cursor, VS Code, LM Studio, BoltAI, and Raycast and configures them for you, then confirms the server starts. No Node.js on your Mac? See the Mac install guide. To set up Claude Code by hand instead: claude mcp add -s user scrivener -- npx scrivener-mcp, then restart it (or run /mcp).

Smithery

npx -y @smithery/cli install scrivener-mcp --client claude

npx, no install

npx scrivener-mcp

or add it to Claude Desktop's config directly:

{
	"mcpServers": {
		"scrivener": { "command": "npx", "args": ["scrivener-mcp"] }
	}
}

From GitHub

npm install -g writerslogic/scrivener-mcp              # latest main; requires pinned Rust/wasm-pack
npm install -g writerslogic/scrivener-mcp#v0.12.0       # a specific release

Homebrew (macOS)

brew install writerslogic/tap/scrivener-mcp

Docker

docker build --platform linux/amd64 -t scrivener-mcp https://github.com/writerslogic/scrivener-mcp.git
docker run --platform linux/amd64 -i --rm -v /path/to/your/projects:/projects scrivener-mcp

The container uses Linux x64 with glibc because the pinned Narrative Lens 0.2.0 package has no Linux ARM64 or musl binary. Linux ARM64 and musl hosts can use this Docker configuration; ARM hosts need Docker's amd64 emulation. The image runs as the non-root node user; mounted project directories must permit that user to read and write.

Any other MCP client: point it at npx scrivener-mcp as a stdio server.

Document management, deterministic analysis, local passage search, and project memory work with no API key. Provider-assisted interpretation, generation and enhancement use Anthropic, OpenAI or OpenRouter. Set more than one and Claude handles chat by default; override with AI_PROVIDER=openai or AI_PROVIDER=openrouter (OpenRouter defaults to anthropic/claude-sonnet-4.6, change it with OPENROUTER_MODEL). If the active provider fails on an account-level error — bad key, no credit, an outage — the server retries on the next one you've configured. If your MCP client supports sampling, chat-based features can run through the client's own model on legacy MCP connections, no separate key needed. MCP 2026-07-28 connections currently use configured providers for AI features; modern multi-round sampling is not implemented. Passage retrieval runs locally through Holographic Memory 0.6.4; interpretation and explanations are optional provider features.

Keys are picked up automatically from:

• ANTHROPIC_API_KEY / OPENAI_API_KEY / OPENROUTER_API_KEY
• ~/.env or ~/.scrivener-mcp/.env
• ~/.anthropic/key, ~/.openai/key, ~/.openrouter/key
• macOS Keychain (anthropic-api-key / openai-api-key / openrouter-api-key)

security add-generic-password -s anthropic-api-key -a anthropic -w sk-ant-your-key-here   # Keychain
export ANTHROPIC_API_KEY="sk-ant-..."                                                      # or export it directly

What You Can Do

For manuscript operations, open a project first. The server has no link to the Scrivener app itself and can't see what's open there. Say "Open my Scrivener project at ~/Documents/My Novel.scriv", or "Discover my Scrivener projects" if you don't know the path. On macOS, "Use the project I have open in Scrivener" works too (the first time, macOS will ask permission to control Scrivener). Do this once per conversation. If the project is also open and unsaved in Scrivener itself, save or close it there first, or the two can write over each other.

Manage your manuscript. Read chapters, create scenes, reorganize the binder, update synopses — all through conversation. "Create a new scene called 'The Reveal' after Chapter 5, and move the old epilogue to the trash."

Analyze your writing. Readability, pacing, style, dialogue quality, emotional arc — grounded in your actual prose, not generic advice. Ask whether a chapter's pacing is off and get specifics: which paragraphs stall, how the scene compares to your other chapters, filter-word density against your own average.

Enhance your prose. Targeted edits: cut filter words, strengthen verbs, vary sentence structure, add sensory detail, turn telling into showing, tighten dialogue, fix pacing.

Track characters and plot. Character profiles, plot threads, and style guides persist with the project across sessions. Save a profile for a character once; a consistency check months later catches contradictions — dialogue that doesn't sound like them, a limp that disappears for a chapter.

Search by meaning. Warm 0.6.4 combines native BM25 with local embeddings and optional cross-encoder reranking. Configure local ONNX models for general semantic search; the default encoder covers a disclosed English story-concept vocabulary. Results include exact passages, locations and saved revisions; changed passages update incrementally. No embedding provider is required. See local search configuration.

Track relationships. Record and query how characters, locations, themes, and plot threads connect. Relationships persist in the project's local SQLite state. Discovery proposes unrecorded character/location pairs using saved-text co-occurrence with source revisions.

Compile and export. Assemble chapters into one manuscript with your own formatting and structure preserved. Export inline as Markdown, HTML, or JSON, or write a DOCX, EPUB, or PDF to disk.

Sign and validate exports. Activate credentials with use_skill({"name":"credentials"}), then call c2pa_capabilities to inspect the installed SDK. Export a PDF with export_project, sign it to a new embedded output with sign_content_credentials, and check that output with validate_content_credentials. Credential operations run locally without an open project and use administrator-configured signing material. They do not establish human authorship or formal C2PA conformance. See Content Credentials for setup, mode coverage and limits.

All Tools

82 tools organized by workflow. To keep token usage low, tools load progressively — project tools at startup, document and search tools once a project is open, the rest on demand. Set SCRIVENER_MCP_EAGER_TOOLS=1 to load everything up front.

Manuscript comparisons include comparisonBaseline.name and its original envelope, alongside the current result's envelope, so both saved source revisions remain identifiable after a baseline name is reused.

Enhancement types: eliminate-filter-words, strengthen-verbs, vary-sentences, add-sensory-details, show-dont-tell, improve-flow, enhance-descriptions, strengthen-dialogue, fix-pacing, expand, condense, rewrite

Intelligent compilation returns provider warnings, explicit quality availability, and saved-source revisions. Rewrites require the complete source to fit 12000 UTF-16 characters; otherwise the complete original is preserved. Quality opinions cover at most 4000 characters and identify that coverage. Marketing generation reads the saved Draft by default; use documentIds to select a smaller scope within the same 12000-character limit. Both reject source changes during generation and provide read_result references for large output.

Activate the credentials skill on demand; these tools do not require an open project.

Format declarations are not proof of native compatibility. Use the Content Credentials guide and per-format native evidence before making coverage claims.

Project memory travels within each .scriv project. Explicit series memberships and shared facts use the server's persistent registry; see editorial workflows.

remember validates nested records and acknowledges the saved snapshot with its revision and record ID. Use that ID for updates; identical creations reuse their ID, and unknown update IDs fail. Unrecorded profile fields, style preferences and writing statistics stay absent. worldBuilding stores named records, for example { name: "Moon Harbor", type: "location" }; all merges custom JSON by key. Requests are capped at 64 KiB and complete project memory at 16 MiB. Large recall results use read_result. Earlier custom world data remains available under recall with all.

Relationship tools use project-owned SQLite without an external service. Storage admits up to 10000 relationships; names contain at most 128 UTF-16 characters and labels 64. Discovery covers the characters and locations already recorded in that graph, up to 256 entities, 1000 Draft documents, 2 Mi UTF-16 source characters and a source-character count multiplied by entity count of at most 134217728. Its score is the fraction of documents mentioning both entities among those mentioning either; it is a candidate association, not a confirmed story fact. Result limits disclose omitted candidates, and large responses use read_result. Relationship queries page in stable identity order using nextOffset and graphRevision; edits require restarting the query. Document cross-reference tools use the character/location registries and saved text, including Research and Draft items excluded from Compile. Document queries read only the requested document; entity-wide queries and suggestions inspect all active text within the same source limits. They disclose missing registries, reject source/registry changes, and preserve every admitted match offset (up to 1000000 positions and 16 MiB of serialized mention records).

Queued analysis always runs local prose analysis and includes readability by default. Optional named entities require a native build with ONNX support and a loaded NER model; unavailable results are explicit. Optional sentiment uses a local lexicon heuristic. Source revisions hash the supplied UTF-8 text, and the envelope identifies this raw policy separately from saved RTF analysis and saved project controls. Project batches preserve input order and run sequentially unless parallel is selected, with at most five documents in flight. Large completed results return a _ref retrieved through read_result. Redis/KeyDB cannot remove an active job; embedded cancellation prevents work that has not started and discards the result of work already running.

suggest_improvements accepts async: true for the same suggestion result through get_job_status. With useContext: true, documentId is required: generation uses up to the first 16000 UTF-16 characters of that saved document and rejects a changed revision. Its result reports the exact excerpt range, truncation and source revision. The default useContext: false sends only the prompt. Suggestions use the configured provider or client sampling, and queued generation is not automatically retried. Large synchronous and queued results use read_result references.

analyze_writing_style accepts 1–64 nonempty text samples, each up to 64000 UTF-16 characters. Its provider prompt includes at most 8000 characters including separators; coverage reports the range used from every sample, including any omitted samples. Missing or malformed profile fields are explicitly unavailable, partial finding lists preserve valid entries, and sample revisions hash the original supplied UTF-8 text. Large profiles also use read_result references.

Guides

• Getting Started -- Installation, configuration, your first session
• MCP Client Setup -- Copy-paste config for Claude Desktop, Claude Code, Cursor, and VS Code
• Writing with AI -- Analysis workflows, enhancement strategies, memory management
• Troubleshooting -- Common issues and fixes
• Token Optimization -- How the server minimizes context window usage
• Architecture -- How the server works, module structure, data flow
• Scrivener Compatibility -- Supported Scrivener versions, platforms, and format coverage
• Content Credentials -- Local C2PA signing and validation, administrator setup, tested modes and conformance limits
• Scrivener File Format -- The reverse-engineered .scriv format, what we read vs. infer, and safe-modification guidance
• Fuzzing -- Jazzer.js target and OSS-Fuzz integration details
• Contributing -- Development setup, code conventions, adding new tools

Requirements

• Stable Node.js ^22.23.4 || ^24.22.1 || >=26.12.1 (24.22.1 or newer on Node 24 recommended)
• Scrivener 3 project files (.scriv)
• Native npm/npx installs: macOS x64/ARM64, Windows x64, or Linux x64 with glibc. Architecture refers to the Node.js runtime.
• Optional: Anthropic, OpenAI, or OpenRouter API key for provider-backed AI features
• Optional: Neo4j for persistence and advanced graph queries; core relationship tools work without it

The patched floors protect Node TLS paths used by the pinned CAI native SDK; SQLite 13 also requires Node-API 10. Node 23 and 25 are excluded. See the runtime compatibility policy.

Development

See the compatibility and upgrade policy for runtime checks, protocol contracts, dependency upgrades and the stable release gate.

git clone https://github.com/writerslogic/scrivener-mcp.git
cd scrivener-mcp
npm install --ignore-scripts
# Install the pinned Rust/wasm-pack tools listed in native/c2pa-binding/README.md
npm run build:c2pa   # Generate the engine from source
npm run dev          # Development mode with hot reload
npm run build        # Compile TypeScript
npm test             # Run tests
npm run typecheck    # Type checking only

Why This One?

A few Scrivener MCP servers exist. Our column describes the current server; other feature claims come from each project's own docs, published package, and advertised tools, last re-read on 2026-10-22; stars, forks, activity, and published version were refreshed 2026-11-02. "No" means undocumented — not

necessarily impossible through the connected AI client.

Documentation truncated — see the full README on GitHub.