Burn — Personal Knowledge Base MCP Server is an MCP (Model Context Protocol)-based server that makes reading data from the Burn read-later app available as an AI-accessible knowledge base. It provides 26 tools and works with Claude, Cursor, Windsurf, and any agent that supports MCP.
Purpose Burn is a read-later app with a 24-hour deadline: saved links automatically "burn" if they are not read or rescued. Its state flow includes: Flame (new links, read within 24 hours or they burn), Spark (read, kept for 30 days), Vault (a permanently kept curated knowledge base), Ash (expired). This MCP server lets AI agents search, triage, organize, and analyze all saved content, aimed at Claude users who want to use their reading history as context.
Main capabilities The server includes 26 built-in tools covering the following categories:
- Triage: the agent acts as a filter, checking the Flame inbox, reading content, and deciding what to keep or burn.
- Collections: the agent acts as a curator, creating collections from saved content.
- Analytics: the agent acts as an analyst, finding patterns and discovering forgotten treasures.
- Auto Feed: the agent acts as a scout and can add monitored sources so that new articles enter Flame. Specific example prompts include: "What have I saved about system design?" "Triage my Flame, which ones should I keep?" "Create a collection from my saved AI bookmarks," and so on.
How to use
- Get a token: create a free Burn account (iOS app or web), open Settings → MCP Server → Generate Token and copy it. The server exchanges the token for a short-lived Supabase session and caches it in ~/.burn/mcp-session.json, so there is no need to log in again on every run.
- Remote method (recommended): use the Streamable HTTP hosted endpoint https://burn-mcp-server.vercel.app/api/mcp , open Settings → Connectors → Add custom MCP in claude.ai, fill in the endpoint and add the same Authorization request header.
- Local method: run the npm package in Claude Desktop, Cursor, Windsurf, or Claude Code (npx burn-mcp-server, stdio transport).
- Start asking questions: natural language prompts are supported, such as "Search everything in my Vault about context engineering and summarize the main schools of thought" and "Create a collection named 'MCP in Production', then write an overview of the agreements and disagreements among my sources."
Environment variables One of BURN_MCP_TOKEN or BURN_SUPABASE_TOKEN is required.
Use cases
- Personal knowledge management: agents search reading history to answer questions, discover patterns, and forgotten treasures.
- Research workflows: create collections on explored topics, and agents write comprehensive overviews.
- Reading triage: agents review the Flame inbox, read content, decide what to keep or discard, solving the problem of read-later app save pileups.
- Cross-tool intelligence: work with Claude Code, Cursor, or Windsurf to make bookmarks context for coding, writing, and thinking.
- Write with your own sources: pull relevant bookmarks and list arguments supported by sources.
Limitations and notes (publicly stated honest boundaries)
- No save/add URL tool: agents can search, read, triage, and organize existing content in Flame/Spark/Vault, but cannot add new URLs to the inbox on their own. You must first save from the iOS app, Chrome web clipper, or burn451.cloud.
- Keyword search, not semantic/vector search: search_vault and search_sparks match titles, tags, and AI-generated key points, not embedding similarity. Search by topic keywords, not vague descriptions.
- No delete function: tools only move bookmarks between Flame → Spark → Vault → Ash, and will not permanently delete any rows via MCP.
- Both remote and local transports are supported, but there is currently no OAuth flow: both methods require the user to generate and paste a token.
- Reading/triage is the mature path; agent write-back is newer, write_bookmark_analysis and update_collection_overview allow agents to write structured notes back to the data, but these write tools are not as thoroughly tested as the read/search/triage tools.
Other information
- Tokens are limited to user data (row-level security), the state flow enforces Flame → Spark → Vault or → Ash, the rate limit is 30 calls per minute per session, and tokens expire after 30 days.
- The free version can save 5 links per day and includes all 26 MCP tools; the Pro version is $4.99 per month for unlimited saves.
- Related links: app burn451.cloud; npm package burn-mcp-server; developer docs burn451.cloud/developers.
- Container and Glama publishing: the container builds the TypeScript source, keeps only production dependencies, and runs the stdio server as a non-root user. Authentication is still a runtime requirement, and the image does not contain tokens. Glama publishing requires using the repository Dockerfile, setting CMD to node dist/index.js, and declaring BURN_MCP_TOKEN as a required secret string.