Sponsored by Deepsite.site

Tag

#anthropic

132 results found

Scratchpad Mcp

scratchpad-mcp is an MCP server that gives AI agents persistent, token-efficient storage. It solves a specific waste problem: agents constantly re-read files they've already seen, re-summarize documents they've already processed, and re-load context they've already understood. Every one of those round-trips burns tokens for no new information. This server fixes that with eight tools designed around how agents actually work: Versioned writes. write_file automatically versions every write and keeps the 10 most recent versions per file. Storage is append-only on success and atomic on failure partial writes can't corrupt state. Structured diffs. read_file accepts a since_version parameter and returns a JSON line-diff against that prior version instead of the full content. Agents that have already seen v1 can ask "what changed in v3?" and get a small structured payload they can reason about, not the entire file again. Append-only logs. append_log and read_log give agents an event-stream they can replay. Cursor-based pagination (since_entry + last_entry_id + has_more) means an agent can checkpoint where it left off and resume cheaply. On-demand summaries. summarize_file calls Claude Haiku to summarize files over ~2000 estimated tokens. Summaries are cached per file version, so repeat calls on an unchanged file cost nothing. The threshold is enforced server-side you can't accidentally pay to summarize something small. Per-agent isolation. Every operation is scoped by an agent_id parameter, so one server instance can serve many agents without leaking state between them. Storage limits. 1 MB per file write, 64 KB per log entry, 1000 files / 100k log entries / 100 MB total per agent sane multi-tenant guardrails out of the box. Backed by a single SQLite file (Postgres migration is on the roadmap). All SQL is parameterized, paths are validated against a strict allowlist, and the security model is documented honestly it's safe for one-user-per-process deployments today, and the V2 plan derives agent_id from the caller's API key for true multi-tenancy. Build agents that remember what they've already seen.