Give your agent your email. Not your mailbox.

Mailvec keeps a synced copy of your mailbox on a machine you control, indexes every message and attachment, and lets Claude or any MCP client search it by keyword and by meaning. Agents read only what they ask for; they can't send, reply or delete. Embeddings and OCR run locally.

Attachments, not just emails.

Mailvec extracts the text from PDFs, Word documents, spreadsheets, slides, calendar invites and contact cards, and indexes it with the email. When a search hit came from an attachment, the result says which one. An agent can read the extracted text, or look at a rendered page when the layout matters: tables, forms, signatures.

Scans too.

A scanned PDF has no text to extract. Mailvec's embedder renders each page and transcribes it with a local vision model (qwen2.5vl:7b through Ollama), then indexes the result like any other text. Photos of documents pasted into emails get the same treatment.

A hosted OCR provider is available if you want it faster; it's off unless you turn it on.

Built for hostile mail.

Anyone can send you an email, so Mailvec treats every message as hostile input. In the Docker deployment, all parsing of mail, HTML, PDFs, Office files and images happens in a separate container with no access to your archive, no credentials and no network route out. Only the sync container holds your mail password. Services run as non-root users with dropped privileges and memory limits. And the server's instructions and every tool description tell the AI that mail content is data from the sender, never instructions.

That last part is a framing, not a guarantee. An agent that can read your mail can be shown a malicious email. Connect Mailvec only to clients you'd trust with your whole mailbox.

$ docker compose exec mcp mailvec doctor

Mailvec preflight
=================

Configuration
  ✓ Version             v0.12.1 (cli + mcp in lockstep)
  ✓ Embedding profile   ollama-legacy (ollama, provider ollama) — mxbai-embed-large @1024d, space ollama:mxbai-embed-large:1024
  ✓ Database            /data/archive.sqlite (4.51 GB)
  ✓ Schema              version 14 (current)
  ✓ sqlite-vec          /app/cli/vec0.so
  ✓ Maildir root        /mail/Fastmail

Pipeline
  ✓ Embedding model     mxbai-embed-large (1024d, matches config)
  ✓ Embedding cover     79,446 / 79,446 messages (100.0%, 330,066 chunks)
  ✓ Last indexed        2026-11-10 14:45 (4m ago)
  ✓ Ollama              reachable at http://‹redacted›:11434 (configured model: mxbai-embed-large)
  ✓ Orphan vectors      none
  ✓ Parser              remote parse service at http://parse:3400 answers /up
  ✓ OCR model           mistral-ocr mistral-ocr-4-0 @ https://‹redacted› — configured; no credentials in this process by design. Check it where OCR runs: `docker compose exec embedder mailvec doctor`.

Services (compose)
  ✓ mbsync sync         last successful sync 14s ago
  ✓ Supervisor          compose (no launchd in a container)
  ✓ mbsync stderr       n/a in a container — the sidecar writes stderr to `docker compose logs mbsync`. Sync outcome is covered by the `mbsync sync` check.

External tools
  ✓ mbsync              n/a — IMAP sync runs in the mbsync sidecar container

MCP
  ✓ MCP /health         200 ok (http://127.0.0.1:3333/health)

18 pass, 0 warn, 0 fail
All clear.

Expect a wait. The first sync and embedding pass on a large mailbox can take hours or days. Keyword search works from the start.