Connect via MCP

Anthra speaks the Model Context Protocol over Streamable HTTP at /api/mcp. There are two ways to connect:

OAuth — for Claude.ai, ChatGPT and other clients that add a server by URL. Paste https://anthra.com/api/mcp, sign in to Anthra and approve the request. No key to copy.
API key — for Cursor, Claude Code, Claude Desktop config files and scripts. Send one of your personal access tokens as a Bearer header.

Endpoint

URL:    https://anthra.com/api/mcp
Method: POST (JSON-RPC)
Auth:   OAuth 2.1 (discovered automatically), or
        Authorization: Bearer ant_live_…
Protocol revision: 2025-06-18

Connect with OAuth (Claude.ai, ChatGPT)

In Claude.ai: Settings → Connectors → Add custom connector. In ChatGPT: turn on developer mode, then add a connector under Settings → Connectors. Use this URL:

https://anthra.com/api/mcp

The client opens an Anthra page that names the app and lists what it is asking to do. Sign in if you need to, check the permissions, and choose Allow. Anything the app generates is charged to your credits, exactly as it would be in the studio.

If the client asks for no particular permissions, it gets books:read, books:write, generation:run and account:read. Publishing (publish:write) and reader analytics need the client to request them. Access tokens last an hour and refresh automatically for up to 30 days of inactivity. To cut an app off, open Account → API keys and choose Disconnect under Connected apps.

For client developers: discovery starts from the WWW-Authenticate header on a 401, which points at /.well-known/oauth-protected-resource (also served at /.well-known/oauth-protected-resource/api/mcp). Authorization-server metadata is at /.well-known/oauth-authorization-server. Anthra supports dynamic client registration, the authorization-code grant with PKCE (S256 only), rotating refresh tokens, token revocation and the resource parameter, which must be https://anthra.com/api/mcp. Redirect URIs must be https, http on localhost / 127.0.0.1, or a native app's private-use scheme (e.g. cursor://…, vscode://…), and are matched exactly.

Connect with an API key

Create a key on Account → API keys with the scopes you want, then add it to your client as below.

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or the equivalent on your OS:

{
  "mcpServers": {
    "anthra": {
      "url": "https://anthra.com/api/mcp",
      "headers": {
        "Authorization": "Bearer ant_live_…"
      }
    }
  }
}

Restart Claude Desktop. You should see anthra appear in the tools list with entries like create_book, generate_outline, write_chapter, and publish_book.

Claude Code

claude mcp add --transport http anthra \
  https://anthra.com/api/mcp \
  --header "Authorization: Bearer ant_live_…"

Cursor

In Cursor’s settings → Tools & Integrations → Add MCP Server:

{
  "mcpServers": {
    "anthra": {
      "url": "https://anthra.com/api/mcp",
      "headers": { "Authorization": "Bearer ant_live_…" }
    }
  }
}

What the model can do

The Anthra MCP server is at full parity with Anthra's in-app writing agent — the same toolkit, exposed over MCP. The model will:

List, read, create, update, and delete books, chapters, and collaborators (books:read, books:write).
Drive the planning surface — theme statement, reader promise, comp authors, beat sheet, POV / tense locks, characters with want / fear / lie, inter-character conflicts, premise interview, premise pitches, outline grader, outline variants, genre detect, era research, mood references, outline snapshots.
Manage the writer's bible — characters, plot outline, theme outline, style guide, worldbuilding, glossary, synopsis, lore, setting bible, era research. Read, write, auto-draft, and roll back via revisions.
Generate outlines, write chapters, rewrite with a directive, run continuity / cliché / repetition / style-drift / spoken-readability / sensitivity checks, and synthesize narration (generation:run).
Run the full prose-analysis suite — filter words, adverb density, sensory balance, concrete nouns, POV slippage, tense consistency, sentence rhythm, dialogue tags, body language, semantic clichés, telegraphing, and repetition — then apply the fixes (cadence variations, show-don't-tell rewrites, verb upgrades).
Convene the specialist / critique panel — call a named specialist, run a critique round-table, skeptic-verify a claim, synthesize as Editor-in-Chief, run a draft tournament, the long-horizon plot doctor, genre-expert lookups, and the debut-reader read.
Gather reader feedback — a test-reader panel, an emotional-arc map, and hook / ending scoring that reads the book the way a first-timer would.
Maintain worldbuilding & continuity — world rules, timeline / geography / weather / travel-time checks, unit consistency, invented terms & calendars, the foreshadowing ledger, object & location trackers, scene-consistency checks, and continuity diffs.
Sharpen character craft — the want / fear / lie roster and conflict matrix, plus dialogue fingerprints, per-character voice consistency, and character-arc tracking.
Do persona & structural editing — track-changes edits with built-in or saved editor personas, sentence variants, tense / POV flips, voice transplant, cut-by-percent, expand, reverse-outline, scene cards, and a synopsis — each landing as a reviewable suggestion you accept or reject.
Grade with scorecards & eval exemplars, and triage the ideas inbox (capture / organise / status).
Collaborate — manage collaborators, hand out chapter assignments, and read the activity feed for a multi-author book.
Cover & production — choose cover color, generate cover / back cover / spine / chapter images, compose the print-ready wraparound spread, render the manuscript to EPUB / DOCX / PDF, export to Markdown, batch-narrate every chapter, and translate the book into another language.
Chapter restructuring — list, edit, split, merge, reorder chapters and roll back to any saved revision.
List the narration voice catalog (list_voices) to pick a voice before kicking off narration.
Publish books and toggle visibility (publish:write).
Read account usage, background-job status, and reader analytics (account:read, analytics:read).

Tools that mutate or spend tokens are gated behind their scope — a key or connected app without generation:run can still let the model list and read your books but won’t allow generation.

Recommended scopes for a chat-only client

books:read
books:write
generation:run
publish:write

Add account:read and analytics:read if you want the model to reason about spend or reader behaviour.

See the full tool reference for every operation, its arguments, and which scope it requires.