Inbox To Action

inbox-to-action

Published on PyPI and listed on the Official MCP Registry (io.github.tarunlnmiit/inbox-to-action), Glama (deployable release, Quality A), and Smithery (MCPB bundle). Registry manifests (server.json, glama.json) ship in the repo.

One command. Your inbox triaged, summarized, drafted, and turned into tasks — in a single agentic pass.

Install

pip install inbox-to-action          # or: pipx install inbox-to-action

uvx inbox-to-action run --mock # zero-install trial (uv) pip install 'inbox-to-action[mcp]' # + MCP server for Claude Code docker run --rm ghcr.io/tarunlnmiit/inbox-to-action # MCP server (stdio)

Try it with zero setup: inbox-to-action run --mock (bundled sample inbox).

📖 Full documentation → [docs/](docs/README.md) — install · providers · Gmail OAuth · multi-account · integrations · MCP & Skill · config · troubleshooting · testing checklist. Quick version: SETUP.md.

---

Why this exists

Most people process their inbox with four separate tools: an email client to read, a task manager to capture to-dos, a calendar to block time, and (increasingly) an AI summarizer to make sense of long threads. Every message gets handled four times.

inbox-to-action collapses all four into one agentic pass. Run one command and get a unified triage report, drafted replies saved to Gmail, and extracted tasks — without ever leaving the terminal, and without ever sending an email automatically.

🔒 Drafts only — never sends

This tool cannot send email. It requests only the Gmail readonly + compose scopes; there is no send scope and no send API call anywhere in the codebase (enforced by a test). Replies are saved as Gmail drafts for you to review and send.

Email bodies flow into the LLM prompt, so a hostile email could try to steer its own classification or a drafted reply (prompt injection). Because every draft is saved for human review and nothing is ever sent automatically, the worst case is a draft you choose not to send. See SECURITY.md.

🌐 What leaves your machine

inbox-to-action reads your email. Where your email content goes for classification depends on the LLM provider you pick — and the default (`openrouter`) is a cloud provider, so an out-of-the-box run sends your subjects + bodies to a third party.

Want privacy? Use ollama (local) or claude (keyless) so nothing is transmitted to a third party. --telegram / --todoist also push subjects/tasks off-box (opt-in). Full breakdown → [PRIVACY.md](PRIVACY.md).

Note: triage-report.md and tasks.md are written to your working directory and

contain private email content. If you run inside a git repo, add them to .gitignore.

What it does

1. Fetches unread email from Gmail (last 24h by default).
2. Classifies each into action_needed · fyi · newsletter · noise.
3. Summarizes long threads (>500 words) into two lines.
4. Extracts tasks with deadlines → local tasks.md (optional Todoist via --todoist).
5. Drafts replies for action_needed mail → saved as Gmail drafts.
6. Flags emails that need a calendar block.

Final output: a single `triage-report.md` with a section per category, drafted-reply previews, a tasks summary, and a calendar list.

Architecture — the agent loop

The model's own classification of each email drives which tools fire next — the pipeline is not hardcoded. The same tool functions back the CLI agent and the MCP server.

flowchart TD

CLI[main.py · typer] --> AGENT[agent.py · agentic loop] AGENT -->|model picks tools per email| T1[classify_email] AGENT --> T2[summarize_thread] AGENT --> T3[extract_tasks] AGENT --> T4[draft_reply → Gmail draft] AGENT --> T5[flag_for_calendar] T1 & T2 & T3 & T4 & T5 --> LLM[llm_client.py · pluggable providers] LLM --> P1[OpenRouter / Ollama / NIM / OpenAI OpenAI-compatible HTTP] LLM --> P2[Anthropic · official SDK · keyless via ant auth login] AGENT --> REPORT[report.py → triage-report.md]

fetch → for each email: classify ─┬─ action_needed → extract_tasks + draft_reply + flag_calendar ├─ fyi / newsletter / noise → record only └─ (long thread) → summarize → render triage-report.md

Quick start (2 minutes)

git clone https://github.com/tarunlnmiit/inbox-to-action.git && cd inbox-to-action

python3 -m venv .venv && source .venv/bin/activate pip install -e '.[mcp]' # installs the inbox-to-action command cp .env.example .env

This installs an inbox-to-action console command (and the python -m inbox_to_action.mcp_server entry point used by Claude Code / Glama).

Free-first: run end-to-end on zero spend

Pick whichever keyless/free path you like — all run the full pipeline at no cost:

Option A — `claude` CLI (keyless, fastest; uses your Claude Code login):

PROVIDER=claude inbox-to-action run --mock     # no API key; needs `claude` on PATH

Option B — Ollama (truly keyless, fully local):

ollama serve            # in another terminal

ollama pull llama3.1 PROVIDER=ollama inbox-to-action run --mock # uses bundled sample inbox

Option C — OpenRouter free model (free signup key):

# put OPENROUTER_API_KEY in .env (free models, $0 spend)

inbox-to-action run --mock # default PROVIDER=openrouter

Option D — inside Claude Code (keyless, Claude Code is the LLM): see below.

--mock uses the bundled sample inbox so you can see a full report with zero Gmail setup. Drop --mock once you've authorized Gmail. Free OpenRouter models are often rate-limited; the client auto-rotates a fallback list and retries with backoff.

Real inbox

# 1. Create OAuth credentials in Google Cloud Console (Desktop app),

download client_secret.json into the project, then:

inbox-to-action auth # one-time consent (read + compose only) inbox-to-action run --since 24h --no-drafts # safe first pass: report only, no writes inbox-to-action run --since 24h # triage the last day (creates Gmail drafts) inbox-to-action run --since 3d --max 40 --todoist

• --no-drafts — classify, summarize, extract tasks, write the report, but create no Gmail drafts. Recommended for a first run.
• --max N — cap emails per account (default 25) to bound cost/volume.
• Automated no-reply senders (security alerts, notifications) never get a drafted reply — the report notes them instead.

Telegram summary (--telegram)

Push a concise summary to your phone after each run — counts, action-needed subjects (with draft-ready status), extracted tasks, and a link to your Gmail Drafts.

# 1. In Telegram, message @BotFather → /newbot → copy the bot token.

2. Message your new bot once (say "hi"), then open:

https://api.telegram.org/bot /getUpdates → copy "chat":{"id": ...}.

3. Put both in .env:

TELEGRAM_BOT_TOKEN=... TELEGRAM_CHAT_ID=...

inbox-to-action run --since 24h --telegram

Off by default (opt-in flag). A send failure never breaks the run. Privacy: this sends email subjects + extracted tasks to Telegram's servers (into your own chat). It's notification only — it never sends email.

Multiple accounts (Gmail + Google Workspace)

Declare accounts in config.json — one merged report, each email tagged with its account. Personal Gmail and Workspace both use the Gmail path (Workspace may need your admin to allow the OAuth app).

{

"accounts": [ { "id": "personal", "kind": "gmail", "label": "Personal Gmail" }, { "id": "work", "kind": "gmail", "label": "Workspace" } ] }

inbox-to-action auth --account personal   # authorize each account once

inbox-to-action auth --account work inbox-to-action run --since 24h # fetches + triages across all accounts

Multiple personal Gmail accounts can reuse one client_secret.json — each gets its own cached token (~/.config/inbox-to-action/tokens/ .json). With no accounts block, the tool uses a single default Gmail account (backwards compatible).

Use inside Claude Code (keyless)

When run inside Claude Code, Claude Code is the LLM — no provider key needed. Two integration paths ship in this repo:

MCP server

Exposes IO-only tools (fetch_emails, save_gmail_draft, append_tasks, write_report). Claude Code does the classify/summarize/extract/draft reasoning itself and calls these.

# after `pip install -e '.[mcp]'`

claude mcp add inbox-to-action -- python -m inbox_to_action.mcp_server

This is the same stdio server that MCP registries (e.g. Glama) build from the bundled Dockerfile (CMD python -m inbox_to_action.mcp_server).

Skill

Copy skills/inbox-to-action/ into your Claude Code skills directory, then type /inbox-to-action. The skill instructs Claude Code to fetch, reason, draft, and write the report — keyless.

Anthropic (keyless via ant auth login)

The Anthropic provider uses the official SDK with a zero-arg client, so it picks up your ant auth login OAuth profile — no `ANTHROPIC_API_KEY` required:

ant auth login

PROVIDER=anthropic inbox-to-action run --mock # default model: claude-opus-4-8

Configuration

All keys live in .env (.env.example is committed). Switch providers with PROVIDER: openrouter (default) · ollama · nim · openai · anthropic · claude · host.

Configure triage (make it *yours*)

The default buckets are generic — newsletters and job alerts are treated as no-action. Override that with config.json (copy config.example.json). Two layers:

• `rules` — deterministic field → category overrides applied before the LLM (fast, free, exact). First match wins. field ∈ sender | subject | body | any.
• `triage_instructions` — freeform guidance injected into the classifier prompt for nuance the model interprets.

{

"triage_instruct