mail-index

mail-index

[npm version] [npm downloads] [GitHub stars] [GitHub forks] [CI] [License: MIT] [Node] [deps]

mail-index downloads a preview of your whole inbox to your machine, then smartly fetches the full text of the messages that matter as you use it. That local index lets your AI agent run true summarization and recall over your entire mailbox — instead of being trapped behind Gmail's search bar.

It works through a local MCP server, so any agent (Claude, Codex, any MCP client) can query it. Local-first — the default index never leaves your machine. Operators who need an always-on connector can instead deploy the optional, single-tenant remote Worker Deployment in their own Cloudflare account; mail-index operates no hosting service. Read-only by default — it never mutates your mail unless you explicitly opt into archive + label edits (a gmail.modify re-auth) or, on the remote Deployment, drafts (a gmail.compose re-auth). It never sends or deletes mail. Both opt-in scopes can send at the token level, so that guarantee is enforced in code (a send guard on the Gmail runner plus test/drafts.test.ts), not by the scope. See ADR-0007 and ADR-0011.

[Claude answering a vague inbox question through the mail-index MCP server, locally, with zero network calls]
_{Ask a vague question → one local MCP call → ranked, snippet-first answer. Interactive version →}

Status: v1.5 — published. Progressive sync, the correspondence graph, the interest engine, curation, the full MCP surface, and the write-back loops are built and tested — and mail-index is live on npm with a .mcpb bundle. Still in progress: the bundled Option A OAuth client and signed one-click installers. Architecture lives in docs/PLAN.md; start with docs/INSTALL.md; every release is recorded in CHANGELOG.md.

[!TIP] New in v1.5 — the optional remote Deployment. mail-index can now also run as a single-tenant Cloudflare Worker in your own account: always-on, synced on a cron, and reachable by remote agents over authenticated MCP (plus a small A2A surface). It serves the same tool registry as the local server, backed by a D1 storage driver that passes the same FTS ranking conformance suite, and it can push — Trigger rules match new mail as it lands and fire signed webhooks. Long O(N) work becomes a queued Job instead of a command handback. You can seed a fresh Deployment straight from your existing local index with mail-index export → D1 import, so it starts with your history already indexed. The local CLI + stdio path is unchanged and stays the default, and mail-index still operates no hosting service — the Worker is code you deploy to infrastructure you own. See docs/INSTALL-worker.md, docs/WORKER-SEED.md, ADR-0008.

Also new locally: every MCP response carries a freshness block and any stale read now auto-triggers a background sync (previously only the digest composites did).

[!TIP] From v1.4 — opt-in mailbox writes + human-readable labels. mail-index can archive a message and edit its labels directly on Gmail — via the archive / label CLI commands and the archive_message / modify_labels MCP tools, on both the gog and gws adapters. It stays read-only by default: writes are unreachable until you opt in with a gmail.modify grant (mail-index never sends or deletes; the scope could send, so a code guard enforces this, see ADR-0011) — enable per account with mail-index setup --account <email> --enable-writes or the bundled scripts/enable-writes.sh. Labels render as their human names everywhere (the index caches Gmail's label catalogue and resolves Label_3546… → "Expedition Insure" in both directions), and you can pass a friendly label name to label --add/--remove. See ADR-0007.

How it works

1. Progressive sync — metadata for the whole mailbox in minutes; bodies fetched selectively.
2. Graph — contacts, domains, threads; centrality + communities over your human (non-bulk) mail.
3. Interest — an engagement score per contact from read/reply/star/importance signals. A seed for your curation, not an autonomous decision.
4. Curate — you (via your agent, or a CLI wizard) confirm who/what matters; that profile drives which bodies get fetched.
5. Query — your agent searches, traverses the graph, and reads the messages that matter, all locally via MCP.

[mail-index prewarms a local SQLite index with the important messages, then the MCP server serves the agent from it, fetching from Gmail again only on a miss]
_{Important messages prewarm a local index; the MCP reads from it, touching Gmail again only when a needed body isn't there yet. Interactive version →}

Quick start

Requires Node 24+ and a Gmail MailSource adapter — gog (recommended) or gws. Reading Gmail needs a Google OAuth client, and you get one two ways: use the mail-index beta client (skip Google Cloud entirely; request access to join the ~100-user test list) or bring your own Google Cloud client (no cap, no request — we walk you through it). See docs/INSTALL.md §2 and docs/oauth-and-verification.md.

npm install -g mail-index                    # or: pnpm add -g mail-index
mail-index init                              # scaffold the config# …connect a mailbox (own OAuth client, read-only) — see docs/INSTALL.md §2 / agent-install.md…mail-index sync  --account personal --since 6momail-index graph build --account personalmail-index search "that contract we discussed"

(Prefer source? git clone …, pnpm install && pnpm build, then run the bins as node dist/cli/index.js ….)

Add to Claude

The MCP server registers in one step, but it still needs a mailbox connected first (see the walkthrough). A .mcpb/claude mcp add only adds the server — it doesn't install the adapter, sign you in, or sync.

• Claude Code: claude mcp add --transport stdio mail-index -- npx -y -p mail-index mail-index-mcp (on Windows, prefix with cmd /c: … -- cmd /c npx -y -p mail-index mail-index-mcp — bare npx/.cmd won't spawn. See docs/INSTALL.md §7.)
• Any MCP client (manual):
  jsonc

  { "mcpServers": { "mail-index": { "command": "mail-index-mcp" } } }

• Claude Desktop: download the .mcpb bundle and double-click it (unsigned during beta — you may need to allow it in System Settings / Windows SmartScreen). The bundle is self-contained (ships its own dependencies, run by Claude Desktop's bundled Node — no npx, network, or system Node needed) so it installs identically on Windows/macOS/Linux. A signed all-in-one installer that also installs the adapter, signs you in, and syncs is still in progress; there is no claude:// install link.

Teach your agent the common moves. The MCP server already tells the agent when to reach for it (purchases, receipts, bookings, "who said what", "catch me up"), and the repo ships a Claude Agent Skill with the typical recipes (find purchases, catch up, find a contact's mail, summarize a sender). Drop it in for Claude Code/Desktop:

mkdir -p ~/.claude/skills && cp -R skills/mail-index ~/.claude/skills/

Full walkthrough (auth, curation, enrichment, scheduled sync, desktop-app gotchas) → docs/INSTALL.md. Driving setup with an agent → docs/agent-install.md.

What to expect — time & storage

The index keeps metadata for every message and full text only where it earns it, so it grows with message count, not mailbox size — about 1.5% of your Gmail. First sync runs ~50 messages/min (one-time, incremental after); search is instant. Start with --since 1mo for value in minutes, then expand. Sizing table & growth path → docs/INSTALL.md §9.

Why it's lighter than Gmail search

Stock Gmail-API MCPs are query-based lookup tools — exact query, a network round-trip per call, raw payloads dumped into the model's context. mail-index answers vague questions from a local recall index. Across a 100-question suite on a real mailbox it answered every question for 15× fewer tokens — and the gap widens exactly where a query-based MCP has no primitive at all: summarizing, relationships, and commitments.

[Token savings by category across 100 inbox questions, mail-index vs a stock Gmail-API MCP: Retrieval 3.0×, Logistics 4.9×, Scheduling 9.2×, Commitments 9.6×, Finance 13.0×, Account 20.3×, Relationship 24.7×, Summarize 32.8× — 15.3× overall.]

Per-category tables, the read-one-message comparison, the tool-by-tool landscape, and how to reproduce it all → docs/COMPARISON.md.

Stack

TypeScript · node:sqlite (no native deps) · SQLite FTS5 · Graphology · @modelcontextprotocol/sdk. Node 24+. Pluggable MailSource adapters; ships two Gmail transports — gog (recommended) and Google's gws.

CLI

Two bins ship: mail-index (CLI) and mail-index-mcp (the stdio MCP server).

mail-index init                          Scaffold the operator config + data dirmail-index sync    --account <a> [--since 30d|1mo] [--all] [--query <q>] [--limit N]mail-index sync    --all-accounts        Sync every account by its policy presetsmail-index enrich  --account <a> [--profile | --rule direct|all] [--sender <s>] [--match <fts>] [--limit N]mail-index graph   build [--account <a> | --all-accounts]mail-index curate  [--account <a>]       Interactive curation wizard (no-agent fallback)mail-index compact [--account <a>] [--now]   Demote summarized bulk bodies (ADR-0003)mail-index search  <terms> [--account <a>] [--limit N] [--enrich]mail-index show    <account:message-id>  Print a message (auto-enriches a meta row)mail-index open    <account:message-id>  Print the provider web URL (no fetch)mail-index archive <account:message-id>  Archive (drop INBOX) — opt-in write; needs gmail.modifymail-index label   <account:message-id> [--add <l>]... [--remove <l>]...  Opt-in write; needs gmail.modifymail-index status  [--json]              Per-account freshness + counts

Writes are off by default. Enable archive + label edits for an account with mail-index setup --account <email> --enable-writes (or the bundled scripts/enable-writes.sh <email>), a gmail.modify grant. On the remote Deployment, create_draft is a third opt-in writer, enabled with a separate gmail.compose grant (&drafts=1). Both scopes can send at the token level; mail-index never sends or deletes because the Gmail runner refuses send paths and test/drafts.test.ts fails CI on any new one. See ADR-0007 and ADR-0011.

Documentation

• docs/INSTALL.md — generic onboarding (install, authenticate a MailSource, init, sync, curate, enrich, add the MCP server, scheduled-sync snippet).
• docs/MCP.md — the 18-tool MCP reference for agent integrators: args, compact result shapes, the index_as_of freshness + command-handback contracts.
• docs/ADAPTERS.md — the MailSource contract and how to write + contract-test a new adapter (gws, gog, gmail-rest).
• docs/INSTALL-worker.md — deploy the optional single-tenant remote Worker to your own Cloudflare account: D1 / Queue / KV resources, secrets, the operator allowlist, connecting an agent, and a verification checklist.
• docs/STAFF-SENT-EVIDENCE.md — disabled private fresh SENT-evidence service, caller/grant fences, and activation prerequisites.
• docs/WORKER-SEED.md — seed a fresh Deployment's D1 from an existing local index (mail-index export → import), instead of re-syncing the whole mailbox from Gmail.
• docs/PLAN-worker.md — the remote-Deployment design: locked decisions, milestones, and the local↔remote parity constraints.
• CHANGELOG.md — what changed in every released version.
• docs/PLAN.md — architecture, data model, and the key decisions (ADR digest).
• SECURITY.md + docs/THREAT-MODEL.md — privacy posture, trust boundaries, prompt-injection stance, and a "verify our claims yourself" runbook. The local-only promise is enforced in CI by an egress guard test — the core (src/) makes no network calls; the one exception is the opt-out launch-shim self-updater (bin/selfupdate.mjs, throttled npm-version check, MAIL_INDEX_NO_AUTOUPDATE=1 to disable), audited by the same guard.

About

Built by Unsold Group — a travel & insurtech company building the groundwork to operate as an AI-native business: local-first, agent-native infrastructure that gives AI real, queryable context to work from. mail-index is one piece of that — giving agents durable memory of a mailbox without handing them the keys to it.

More: unsold.group/al

Project growth

[Star History Chart]

Feedback & contact

There's no telemetry in the local install. A self-hosted remote Deployment sends tool-usage metadata (never arguments or results) only if its operator sets POSTHOG_PROJECT_TOKEN (INSTALL-worker.md). We only know what you tell us, so feedback is genuinely welcome:

• 💬 Discussions — questions, ideas, how you use it
• 🐞 Report a bug · 💡 Request a feature
• 🔒 Security/privacy issues → SECURITY.md (private)
• 🌐 unsold.group/al

Inside your agent you can also just say "report a mail-index bug" — the MCP server points the agent to a GitHub link for you to submit (it never sends anything itself). See SUPPORT.md.

License

MIT