mboxer

Create NotebookLM-ready Markdown source packs from Gmail MBOX exports, with local SQLite storage, JSONL exports, and CSV/JSON manifests for search, RAG, and archive review.

mboxer is a local-first email archive processor designed around a common problem:

You can export Gmail as an MBOX file,
but a raw MBOX archive is not useful for NotebookLM, RAG, review, or analysis.

mboxer turns that raw archive into organized, structured, reusable knowledge assets.

Gmail / Google Takeout
  → MBOX file
  → local SQLite index
  → organized Markdown source packs
  → NotebookLM, RAG, search, review, JSONL, and future tools

Why this exists

Gmail archives often contain years of valuable personal, professional, legal, financial, operational, project, and organizational history.

Google Takeout makes it possible to export that history as an .mbox file, but the exported file is not immediately useful for modern AI workflows.

NotebookLM works best with readable, focused, well-organized source documents.

RAG systems work best with structured, chunkable records.

Spreadsheets work best with clean rows and metadata.

Local review works best when everything is inspectable before anything is uploaded.

mboxer bridges that gap.

Primary use case: Gmail MBOX to NotebookLM

The main selling point of mboxer is converting Gmail MBOX exports into clean, category-organized Markdown files that can be used as NotebookLM sources.

Instead of uploading one giant raw archive, mboxer creates structured source packs like:

exports/notebooklm/
  primary-gmail/                 # output is nested under the account key
    finance/
      invoices/
        2024/
          finance-invoices-2024-001.md
    legal/
      contracts/
        2023-2024/
          legal-contracts-2023-2024-001.md
    projects/
      product-launch/
        2026/
          projects-product-launch-2026-001.md
    operations/
      vendor-correspondence/
        2025/
          operations-vendor-correspondence-2025-001.md

The goal is to make exported Gmail content easier to:

• upload into NotebookLM
• organize by topic or category
• review before upload
• split into useful source packs
• preserve context from email threads
• exclude sensitive or irrelevant material
• reuse later for RAG, search, or analysis

What mboxer produces

NotebookLM Markdown source packs

Markdown is the primary output format.

Each exported file preserves useful email context:

• subject
• sender, date, and message ID
• category and source account in the pack header
• body text after the selected export projection

Recipient lists and attachment references remain in SQLite; the current NotebookLM renderer does not include them in Markdown packs.

Export output is split by category, year, and size band to respect NotebookLM source limits. A CSV manifest (manifest.csv) and JSON manifest (manifest.json) are written under <out>/<account-key>/ for each export run. Re-export replaces the account's managed generation and removes obsolete packs listed in its prior manifest. Packing counts complete rendered files and fails before publication if the configured source budget cannot hold the full export. See limits and publication behavior for ownership, recovery, and temporary disk requirements.

SQLite database

SQLite is the durable local project index.

The schema tracks:

• accounts and MBOX sources
• messages with normalized metadata and body text
• Gmail label associations
• thread groupings with participant and date ranges
• ingest runs with resumable checkpoint keys
• ingest errors per run
• attachments with SHA-256, content type, and extraction status
• classifications per message and per thread
• category taxonomy with locked/global flags
• category proposals for review and approval
• export items and export run records
• security findings per message

JSONL exports

JSONL is intended for RAG pipelines, embeddings, local LLM tools, and structured downstream processing.

Each line represents one message with clean body text, metadata, and classification context. Account key is injected into the output path automatically to keep multi-account exports separated, and a <name>.manifest.json is written alongside the JSONL file.

Other output formats

Row-per-message CSV export and external API delivery are not implemented. CSV manifests describe exported source packs; they are not message-data exports.

Current implementation status

The core pipeline is implemented and working.

Implemented:

• MBOX ingest into SQLite using Python's mailbox stdlib
• resumable ingest with per-run checkpoint keys and batch commits
• deduplication via INSERT OR IGNORE on message identity
• multi-account separation with per-account keyed storage
• message normalization: subject, sender, recipients, dates, body text, body hash, word count
• Gmail label parsing and storage
• thread grouping with participant aggregation and date ranges
• attachment extraction to disk with SHA-256 and content-type tracking
• rule-based classification at both message and thread level
• thread-level rule classification with message inheritance
• assign (confidence 1.0) and assign_hint (confidence 0.75) rule actions
• category taxonomy with locked categories and proposal workflow
• category review, approval, and rejection via CLI
• security scan and scrub hooks
• five export content profiles: raw, reviewed, scrubbed, metadata-only, exclude
• residual-findings export gate (allow / warn / block) that re-scans projected export text
• NotebookLM Markdown export with category directories, year bands, and size-limit profiles
• export dry-run mode
• JSONL export
• CSV + JSON export manifests with provenance/lineage fields
• five NotebookLM limit profiles: standard, plus, pro, ultra, ultra_safe
• CLI with subcommands for all pipeline stages
• YAML config loading with deep (dotted) key access
• pyproject.toml packaging with an optional dev extra; version derived from git tags via setuptools-scm

Not implemented: LLM classification, automatic category proposals, a web UI, row-per-message CSV, and external delivery. The application currently uses explicit local commands for ingest, classification, scanning, and export.

Project identity

Project name:      mboxer
Python package:    mboxer
CLI command:       mboxer
Default database:  var/mboxer.sqlite
Entry point:       mboxer.cli:main
Module entry:      python -m mboxer
Python requires:   >=3.11 (tested on 3.11 and 3.12)
Versioning:        git tags via setuptools-scm

Implementation

PROJECT.md maps the source modules, shared components, and tests. Architecture describes data flow and storage boundaries.

Quick start

python -m venv .venv
source .venv/bin/activate

pip install -e .

mboxer --help

Copy and customize the example config:

mkdir -p config
mboxer config-example > config/mboxer.yaml

Without --config, mboxer loads a legacy local config/mboxer.example.yaml if present, otherwise the bundled defaults. This works from any directory after installation. Pass --config config/mboxer.yaml to use your edited configuration.

First run

Complete walkthrough from a fresh checkout to a dry-run export.

1. Initialize the database

mboxer init-db --config config/mboxer.yaml

2. Register your account

mboxer account add primary-gmail \
  --display-name "Primary Gmail" \
  --email user@example.com \
  --config config/mboxer.yaml

3. Verify the account was registered

mboxer account list --config config/mboxer.yaml

4. Ingest a small test archive first (see warning below)

mboxer ingest data/mboxes/primary-gmail/sample.mbox \
  --config config/mboxer.yaml \
  --account primary-gmail \
  --source-name "Sample" \
  --extract-attachments \
  --resume

5. Classify with rules

mboxer classify \
  --config config/mboxer.yaml \
  --account primary-gmail

6. Review categories

mboxer review-categories \
  --config config/mboxer.yaml \
  --account primary-gmail

7. Run a security scan

mboxer security-scan \
  --config config/mboxer.yaml \
  --account primary-gmail

8. Dry-run export to verify output shape

mboxer export notebooklm \
  --config config/mboxer.yaml \
  --account primary-gmail \
  --profile ultra_safe \
  --dry-run

9. Real export when ready

mboxer export notebooklm \
  --config config/mboxer.yaml \
  --account primary-gmail \
  --profile ultra_safe \
  --out exports/notebooklm

Warning: test with a small MBOX before ingesting large archives.

Gmail MBOX exports can exceed several gigabytes for long-lived accounts. Before ingesting a full archive:

1. Extract a small slice of messages into a separate .mbox file and ingest that first.
2. Run mboxer export notebooklm --dry-run to verify the output shape.
3. Review the generated exports locally before uploading anything to a cloud service.

--resume makes ingest restartable, but a full ingest of a large archive still takes significant time and disk space. Export --dry-run performs the same projection and packing work using temporary staging, without publishing output or recording export rows.

Getting a Gmail MBOX file

You can export Gmail data from Google Takeout / Google Data Request.

The typical flow is:

1. Request an export of your Gmail data.
2. Download the archive from Google.
3. Extract the downloaded archive locally.
4. Locate the .mbox file.
5. Ingest the .mbox file with mboxer.
6. Export organized Markdown files for NotebookLM.

Example:

mboxer ingest data/mboxes/archive.mbox \
  --config config/mboxer.yaml \
  --source-name "Primary Gmail Archive" \
  --account primary-gmail \
  --extract-attachments \
  --resume

Intended workflow

mboxer ingest data/mboxes/archive.mbox \
  --config config/mboxer.yaml \
  --source-name "Primary Gmail Archive" \
  --account primary-gmail \
  --extract-attachments \
  --resume

mboxer classify \
  --config config/mboxer.yaml \
  --account primary-gmail \
  --level thread

mboxer review-categories \
  --config config/mboxer.yaml \
  --account primary-gmail

mboxer security-scan \
  --config config/mboxer.yaml \
  --account primary-gmail

mboxer export notebooklm \
  --config config/mboxer.yaml \
  --account primary-gmail \
  --profile ultra_safe \
  --out exports/notebooklm

mboxer export jsonl \
  --config config/mboxer.yaml \
  --account primary-gmail \
  --out exports/rag/messages.jsonl

Configuration and global flags

Runtime commands accept two common flags (config-example only prints the bundled YAML):

• --config PATH — path to your YAML config. An explicit missing or invalid file is an error. If omitted, a local config/mboxer.example.yaml takes precedence over bundled defaults. A config/mboxer.yaml file is used only when explicitly selected.
• --db PATH — override the SQLite database path. Otherwise the path comes from paths.database (then project.default_database) in config, defaulting to var/mboxer.sqlite.

Print the canonical example with mboxer config-example; its source is src/mboxer/defaults.yaml. It covers ingest batch size, classification rules, locked taxonomy, security/redaction policy, NotebookLM limit profiles, and JSONL options. Unimplemented placeholder settings have been removed from the example; old copies remain ignored. Ingest resume and attachment extraction use CLI flags. NotebookLM format and split_strategy settings are descriptive manifest metadata, not configurable behavior. There is no environment-variable configuration support.

Account commands

mboxer account add <key> --display-name "..." --email you@example.com [--provider gmail] [--notes "..."]
mboxer account list
mboxer account show <key>
mboxer account update <key> [--display-name "..."] [--email ...] [--notes "..."]

When exactly one account exists it is auto-selected (with a notice). When more than one exists, account-scoped commands require --account <key> (or --accounts key1,key2 for a combined NotebookLM export).

Useful ingest flags

• --resume — restart an interrupted ingest from its last checkpoint.
• --extract-attachments — extract attachments to data/attachments/ with SHA-256 + content type.
• --create-account — create the --account key on the fly if it does not exist yet.
• --force — reprocess messages even if already present.

Multi-account support

mboxer supports multiple separate Gmail accounts and archives in the same local project.

Each ingested source is tracked by account, source name, import run, and original MBOX file.

Example account keys:

primary-account
work-account
business-archive
organization-archive
project-archive

To export multiple accounts into a single NotebookLM run:

mboxer export notebooklm \
  --config config/mboxer.yaml \
  --accounts primary-account,work-account \
  --profile ultra_safe \
  --out exports/notebooklm

NotebookLM source-pack strategy

NotebookLM exports are Markdown-first and organized by category directories.

Filenames remain meaningful even if the folder hierarchy is flattened during upload.

category-topic-year-sequence.md

Examples:

finance-invoices-2024-001.md
legal-contracts-2023-2024-001.md
projects-product-launch-2026-001.md
operations-vendor-correspondence-2025-001.md
research-literature-review-2024-001.md
support-customer-requests-2025-001.md

Profiles: two independent controls

mboxer has two different "profile" settings on export notebooklm, and they do different things:

They are covered separately below.

NotebookLM limit profiles (--profile)

Limit profiles bound how many source files an export produces and how large each one gets. They are defined in the bundled example (mboxer config-example).

Use ultra_safe as the default for large NotebookLM-oriented workflows where you want to preserve headroom for manual sources, attachments, PDFs, and later additions.

Source, word, and byte limits can be overridden on the CLI; max_messages_per_source is configured in YAML:

mboxer export notebooklm \
  --profile ultra_safe \
  --max-sources 400 \
  --target-words 200000

Full set of limit overrides: --max-sources, --reserved-sources, --target-sources, --target-words, --max-words, --target-mb, --max-mb. Two escape hatches relax the built-in guardrails:

• --allow-full-source-budget — allow the export to use the full max_sources budget (ignore reserved_sources headroom).
• --force — override the 200 MB per-source safety ceiling.

Export content profiles (--export-profile)

Content profiles decide how much of each message body actually leaves the database. Every message gets an effective profile from its classification rule (export_profile: in a rule) or, failing that, from security.default_export_profile (the example config uses scrubbed).

--export-profile overrides the effective profile for a whole run, but only accepts raw, reviewed, scrubbed, or metadata-only. exclude is not a CLI choice — it is applied per message through classification rules or config, so that "do not export" stays a governed, per-category decision rather than a global switch.

Residual-findings gate (--findings-policy)

After a record is projected for export, the projected text is scanned again. --findings-policy (default from security.on_residual_findings, which the example config sets to warn) controls what happens if detected-sensitive patterns survive:

• allow — write the export; record residual counts in the manifest.
• warn — write the export, record counts, and print a counts-only warning.
• block — abort before publishing output or export ledger rows; the command exits with status 2. Temporary staging may have been written and is cleaned up.

Classification strategy

Classification uses deterministic rules, with no network calls.

Rules match on sender domain, sender address fragment, and subject keywords. Each rule assigns a category_path, sensitivity, notebooklm_priority, and export_profile.

At thread level, a matching rule is applied to the whole thread and then inherited down to all messages in the thread.

Rules support two assignment modes:

• assign — confident match, confidence 1.0
• assign_hint — soft match, confidence 0.75

Classification can be scoped by account and run at message or thread level. classification.level sets the default (bundled value: thread); --level takes precedence. An invalid configured level is an error before classification:

mboxer classify --level thread --account primary-gmail

Category taxonomy

Categories are slash-delimited paths that become directory hierarchies in exports.

medical
medical/hospital-billing
medical/pharmacy
legal
legal/law-firm-correspondence
finance
household/utilities
postal/usps-informed-delivery
noise/marketing
noise/spam

Locked categories are defined in config and cannot be deleted.

Proposal review and approval are implemented, but rule classification does not generate proposals. Existing pending database proposals appear in review-categories; approval creates or activates their category:

mboxer approve-category <proposal_id>
mboxer reject-category <proposal_id>

Security stance

mboxer assumes mail archives contain sensitive material.

All exporters write local files. A raw export retains sensitive body text; there is no upload or destination enforcement.

The security pipeline:

ingest
  → normalize
  → classify
  → security-scan
  → export projection and scrubbing
  → residual findings check
  → local files

The security config block controls the default content posture, the residual-findings gate, and which redaction passes run before export:

security:
  default_export_profile: scrubbed       # fallback content profile when no rule sets one
  scan_enabled: true
  scrub_enabled: true
  on_residual_findings: warn             # allow | warn | block (maps to --findings-policy)
  # Attachment scanning and quarantine are not implemented.
  redact_email_addresses: true
  redact_phone_numbers: true
  redact_ssn_like_numbers: true
  redact_credit_card_like_numbers: true
  # physical_addresses is a reserved/planned detector name, not active today

Detection today is a deterministic in-process regex registry covering email addresses, phone numbers, SSN-like values, and credit-card-like values. Physical-address, medical, legal, and credential detectors are reserved future names, not active detection or scrubbing.

Cloud-oriented exports should use reviewed, scrubbed, or metadata-only content profiles.

Development

See CONTRIBUTING.md for the branch workflow and verification guide.

pip install -e ".[dev]"
python tests/fixtures/make_synthetic.py
pytest

Linting and type checking:

ruff check src/
mypy src/

CI checks PRs and pushes to dev and master. It combines Ruff/mypy, tests Python 3.11 and 3.12 (coverage on 3.11), and builds/installs a wheel in an isolated environment outside the checkout. Configure the stable CI gate check as required in branch protection; docs-only changes still complete that gate. Superseded runs are cancelled. Randomized test order runs weekly and on demand. Runtime migration tests run with the suite; there is no separate snapshot-only schema check.

To verify a distribution locally:

python -m pip install build
python -m build
python scripts/smoke_wheel.py dist/*.whl

Keep only the wheel being tested in dist/ when using the wildcard. Build artifacts are ignored by Git. Test fixtures are synthetic; regenerate them with:

python tests/fixtures/make_synthetic.py

Design goals

mboxer should be:

• NotebookLM-friendly
• Gmail MBOX-focused
• local-first
• privacy-conscious
• resumable
• inspectable
• useful without a cloud service
• useful with local LLMs
• useful with future RAG systems
• safe for sensitive archives
• flexible enough for multiple Gmail accounts
• structured enough to support future application features

Non-goals

mboxer is not intended to be:

• a Gmail client
• a replacement for Gmail search
• a hosted SaaS product
• a tool that uploads raw email archives by default
• a black-box AI classifier
• a cloud-first archive processor

Troubleshooting / FAQ

Config file not found: config/mboxer.yaml You passed an explicit config path that does not exist. Create its parent directory, run mboxer config-example > config/mboxer.yaml, then edit it; or omit --config to use defaults.

<command> requires --account when multiple accounts exist More than one account is registered. Pass --account <key>; list keys with mboxer account list.

Unknown NotebookLM profile 'x'. Available: ... --profile must be one of standard, plus, pro, ultra, ultra_safe (or a profile you added under exports.notebooklm.profiles).

effective source budget is zero; reduce reserved_sources reserved_sources is greater than or equal to max_sources for the chosen profile. Lower it, or pass --allow-full-source-budget.

max_bytes_per_source exceeds 200 MB safety limit; pass --force to override A per-source byte limit above 200 MB is rejected by default. Pass --force only if you really intend sources that large.

BLOCKED: would export residual detected-sensitive items ... (exit code 2) Your findings policy is block and the projected export text still contains detected-sensitive patterns. Scrub/redact further, move the affected categories to metadata-only/exclude, or rerun with --findings-policy warn (or allow) if that is acceptable for the run.

mboxer: command not found Activate the virtualenv and install the package: source .venv/bin/activate && pip install -e ..

Ingesting a full archive is slow or huge. Expected — Gmail archives can be many GB. Slice a small .mbox first, use --resume, and validate shape with mboxer export notebooklm --dry-run before a real export.

Does classification use an LLM? No. Classification uses deterministic rules. The unused Ollama configuration example and classify --model placeholder have been removed; --model is now rejected instead of silently running rule classification.

Releases

The package version is derived from git tags by setuptools-scm — there is no hand-edited version string in pyproject.toml. The most recent vX.Y.Z tag determines the version of a build.

Publishing is manual and tag/release-driven:

1. Push a semantic-version tag, e.g. git tag v0.2.0 && git push origin v0.2.0.
2. Create a GitHub Release for that tag.
3. Publishing the GitHub Release triggers .github/workflows/publish.yml, which builds the sdist and wheel (python -m build) and uploads them to PyPI using trusted publishing (OIDC, the pypi environment) — no API token is stored in the repo.

There is no automatic patch-bump-on-merge workflow; choosing the next version number is a deliberate step you take when you tag a release.

License

Apache License 2.0. See LICENSE.