About author-mcp-server
author-mcp-server connects MCP-compatible AI agents to Seizn Author’s canon data and writing tools. Fiction writers can recall and approve canon, check prose for conflicts, search project entities, and explore timelines and relationships through the Seizn API.
Key features
- 01Recall canon for named entities
- 02Check prose for canon conflicts
- 03Approve facts as canon
- 04Semantic search across project entities
- 05View chapter timelines and relationship graphs
Use cases
- 01Recall a character’s established details
- 02Check a scene for continuity conflicts
- 03Search project canon and entity relationships
- 04Review chapter beats across a selected range
MCP support
Exposes six tools to MCP-compatible AI agents and calls the Seizn Author REST API using a Bearer token. The server supports stdio, streamable-http, and legacy SSE transports.
Configuration
Local server config is the same for every client. Copy it into your client’s MCP config file.
npx -y @seizn/author-mcp-serverSEIZN_API_KEYSecretBearer token from https://seizn.com/dashboard/account/api-keysSEIZN_API_BASE_URLOverride the API base URL (default: https://seizn.com/api/v1)Local servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.
Tools
6Saved tool list · updated 2w ago
| Parameter | Type | Required | Description |
|---|---|---|---|
| name | string | Yes | Entity name to recall (e.g. character name, place name) |
| projectId | string | Yes | Seizn project id (find in dashboard URL or via list_projects API) |
Details
From the project README@seizn/author-mcp-server
MCP server exposing Seizn Author canon recall, conflict check, semantic search, timeline, and relationship graph tools to any MCP-compatible AI agent (Claude Desktop, Claude Code, Cursor, Cline, Continue, Zed, etc.).
Bring Seizn canon into the AI tool you already use. No new editor required.
What it does
When connected, your AI agent gains six tools:
| Tool | Purpose |
|---|---|
seizn_author_recall | Get canon for a named entity (character, place, object, rule, event, promise) — returns last 3 mentions, current state, pending conflicts |
seizn_author_check | Check a passage of prose for canon conflicts (P1 critical / P2 warning / P3 stylistic) |
seizn_author_remember | Approve a fact as canon (writer-confirmed source of truth) |
seizn_author_search | Semantic search across all project entities |
seizn_author_timeline | Chapter-by-chapter beats, optional range filter |
seizn_author_graph | Relationship graph subset rooted at an entity |
All tools call the Seizn Author REST API (/api/v1/*) with your Bearer token.
Setup
1. Get a Seizn API key
Generate one at https://seizn.com/dashboard/account/api-keys (Free tier: 100 calls/day, BYOK required).
2. Install (npm)
npm install -g @seizn/author-mcp-server
Or use npx:
npx @seizn/author-mcp-server
3. Wire into your MCP client
Claude Desktop
Add to claude_desktop_config.json:
{
"mcpServers": {
"seizn-author": {
"command": "npx",
"args": ["-y", "@seizn/author-mcp-server"],
"env": {
"SEIZN_API_KEY": "sk_seizn_..."
}
}
}
}
Claude Code
claude mcp add seizn-author -e SEIZN_API_KEY=sk_seizn_... -- npx -y @seizn/author-mcp-server
Cursor / Cline / Continue
See your client's MCP server documentation. The package supports:
- stdio (default for desktop clients)
- streamable-http (for hosted scenarios)
- SSE (legacy)
Configuration
| Env var | Default | Description |
|---|---|---|
SEIZN_API_KEY | (required) | Bearer token from seizn.com/dashboard/account/api-keys |
SEIZN_API_BASE_URL | https://seizn.com/api/v1 | Override for staging or self-hosted |
Usage examples
In your AI chat:
'Recall everything Seizn knows about Seoyun in project ch7-draft.' → invokes
seizn_author_recallwith name=Seoyun
'Check this scene for canon conflicts: ...' → invokes
seizn_author_checkwith the pasted text
'Show me the timeline from Ch.5 to Ch.7.' → invokes
seizn_author_timelinewith from=ch5, to=ch7
The AI decides when to call each tool based on the conversation. You don't have to remember tool names.
Pricing
| Plan | Quota | Price |
|---|---|---|
| Free | 100 recall calls/day | $0 |
| Indie | 1,000 calls/month | $9/month |
| Pro | 10,000 calls/month | $19/month |
| Studio | 100,000 calls/month + 5 API keys per user | $99/month |
| Studio Managed (Q3+) | + 500 managed Opus calls/month | $299/month |
| Enterprise | Volume + SOC 2 + SSO | Custom |
Track 2 plan covers Track 2 surface (API + MCP) only. Seizn Author has separate plans for the web dashboard (Track 1, KRW) and the Tauri desktop app (Track 3, KRW). One Stripe customer holds multiple subscriptions, billed independently. See https://seizn.com/pricing for cross-track upsell details.
No subscription required for AI inference. Tools split into two groups:
| Tool | Backend LLM | Needs BYOK? |
|---|---|---|
seizn_author_recall / _remember / _graph / _search | ❌ DB only | ❌ Works on Free, no key needed |
seizn_author_check / _timeline | ✅ Sonnet/Opus | ✅ BYOK or Studio Managed |
The Seizn subscription covers the canon backend and recall infrastructure. The two AI-enhanced tools (check, timeline) require either your own LLM key (BYOK, any tier) or Studio Managed ($299/mo) which includes 500 Opus calls.
Your host AI agent's LLM is separate. Whether you use Claude Pro/Max subscription via Claude Code/Desktop, Cursor Pro, or BYOK in Cline/Continue — the host handles its own LLM cost. Seizn doesn't double-charge.
See https://seizn.com/pricing for full details.
Versioning
While the package is on 0.x.y we may ship breaking changes on minor bumps (0.2.0 etc.) — we'll call them out clearly in CHANGELOG.md. Once we tag 1.0.0, strict semver applies: major = breaking, minor = additive, patch = fix-only.
Each release is published with a matching GitHub tag (v0.1.0, …) and a CHANGELOG.md entry.
License
MIT © Litheon LLC
Issues / contributions
https://github.com/litheonhq/seizn/issues (label: track-2-platform)
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.