What's new

Four new built-in MCP tools on writeable endpoints: ingest_url (fetch + parse a URL), ingest_file (small files inline as base64), and request_upload_url + ingest_uploaded_file (large files via signed PUT). Agents can extend a corpus mid-conversation. Async via the same job queue as the rest of the platform; polling via get_write_job_by_id.

Toggle 'Allow agent writes' on any Pro+ endpoint. Schema-gen produces insert_<table>, update_<table>_by_id, and delete_<table>_by_id tools. Async-by-default via a write_jobs queue; every write is audit-logged. Free stays read-only by design.

OAuth tokens now carry a scope (mcp or mcp:write). The consent screen at shipmcp.io/oauth/authorize shows a separate 'Also grant write access' checkbox whenever you have at least one writeable endpoint — works regardless of which scope the client requested.

Claude.ai web, Cursor, MCP Inspector, and any modern MCP client can connect via standard OAuth (RFC 8414, 7591, 7636 PKCE). No more pasting bearer tokens. Paste your endpoint URL → consent → connected.

Endpoints archive original uploads to R2 by default (toggle per endpoint). Agents can fetch files via MCP resources/list + resources/read, or via the auto-generated get_documents_file tool. Files ≤ 4 MB inline as base64; larger files come back as 120-second signed URLs.

Documents now carry a 'kind' (text / pdf / audio / image / office / web) and per-kind metadata tables (pdf_metadata, audio_metadata, image_metadata, office_metadata, web_metadata). Schema-gen produces typed cross-table tools like get_audio_metadata_for_document automatically.

Audio ingest passes language='en', vad_filter=true, and a punctuation/speaker-turn prompt to Whisper for noticeably better transcripts. Empty-transcript retry without VAD covers instrumental music; clear error message when no speech is detected.

Browser Rendering now uses waitUntil='load' + a 3s post-load action timeout so Notion-hosted sites and other JS-heavy SPAs hydrate before extraction. Empty-extraction guard rejects results <50 chars with an actionable error.

Sign in with a one-time email link from /auth/magic-link. Doubles as the official forgot-password recovery path. Tokens expire in 15 minutes, are single-use, and only the SHA-256 hash hits our database.

Sixteen new file formats via Cloudflare Workers AI's toMarkdown binding. Drop a PDF or a recording — we transcribe and extract a Markdown corpus, then run the standard ingest pipeline.

From any active endpoint, click Add data to upload more rows or pull in a related table. Schema is re-introspected, llms.txt is regenerated, no new endpoint URL.

The endpoint detail page now polls a 7-step tracker every 2 seconds: queued → fetching → provisioning_db → loading_data → introspecting → publishing → active. No more refreshing.

Name your endpoint at upload time. Optional — we still auto-slug from the filename if you skip it.

Newly-minted tokens are blurred by default with dedicated Reveal/Hide and Copy buttons. Same treatment in the settings page and the endpoint detail page.

Markdown corpora, tab-separated values, plain text, and HTML join CSV / JSON / NDJSON / SQL / Excel as native parsers. No Workers AI required.