tabrewind

tabrewind is a local-first tool for finding "lost tabs" and recent pages quickly.

Many browsers already store history in local SQLite databases, which makes retrieval reliable and easy to query. tabrewind builds on that with a search and indexing layer focused on Zen/Firefox first.

what it does

• reads browser history from local SQLite data in read-only mode
• supports direct SQLite/FTS queries for exact and keyword search
• supports vector search with local embeddings
• supports hybrid search (BM25 + vector)
• exports deterministic outputs for repeatable indexing/search

usage modes

• CLI workflow for direct local usage
• agent workflow (for example OpenCode/Codex-style agents) that can run queries and search flows

tech

• Python (for now - may switch to Go/Rust/Zig later)
• SQLite
• sqlite-vec
• SQLite FTS5/BM25

local model runtime

tabrewind uses llama.cpp for local model execution.

• use any model that is supported by your installed llama.cpp build
• download model files locally before running model-backed workflows
• no hosted model API is required for this project

Example model download via llama.cpp:

llama-cli \
  --hf-repo ggml-org/embeddinggemma-300M-qat-q4_0-GGUF \
  --hf-file embeddinggemma-300M-qat-Q4_0.gguf \
  -n 0 -p ""

CLI workflow

Initialize a local config file once:

uv run main.py init

Inspect or update config values:

uv run main.py config show
uv run main.py config set llama.base_url http://127.0.0.1:8080
uv run main.py config set ingest.embedding_workers 4
uv run main.py config set ingest.embedding_batch_size 8

Configure host filtering rules (top-to-bottom, last match wins):

# deny sensitive domains
uv run main.py domains add --deny "gmail.com"
uv run main.py domains add --deny "*.gmail.com"
uv run main.py domains add --deny "*.bank.*"

# allow a specific subdomain after a broad deny
uv run main.py domains add --allow "mail.google.com"

# inspect, reorder, and validate final decision
uv run main.py domains list
uv run main.py domains move --from-index 4 --to-index 2
uv run main.py domains check "https://mail.google.com/mail/u/0/#inbox"

# preview resolved allow/deny view for discovered hosts in recent history
uv run main.py domains preview --since-days 2 --browsers zen

Rules use a leading prefix in config:

• -pattern denies a host glob
• +pattern allows a host glob
• patterns may optionally include a path glob (for example -mail.google.com/mail/*)
• unmatched hosts are allowed by default
• rules are applied top-to-bottom, and the last matching rule wins

You can also set all rules at once via config:

uv run main.py config set ingest.domain_rules -- "-gmail.com,-*.gmail.com,-*.bank.*,+mail.google.com"

Common privacy rule recipes (order matters, last match wins):

# block personal inbox + broad banking surfaces
uv run main.py config set ingest.domain_rules -- "-gmail.com,-*.gmail.com,-*.bank.*,-*.paypal.com"

# allow one work-safe mailbox after a broad deny
uv run main.py config set ingest.domain_rules -- "-*.google.com,+mail.google.com"

# block only sensitive app routes, keep public pages
uv run main.py config set ingest.domain_rules -- "-app.example.com/settings/*,-app.example.com/billing/*"

Tip: run uv run main.py domains preview --since-days 2 --browsers zen to review the final resolved ALLOW/DENY view before ingesting.

Start llama-server in embedding mode. If you plan to send concurrent embedding requests, set --parallel to match or exceed your configured worker count:

llama-server -m <path_to_model.gguf> \
  --embedding \
  --pooling mean \
  --parallel 4 \
  --port 8080

Run ingestion and print vec[:10] previews:

uv run main.py ingest --max-items 20

Run schema-backed ingestion into tabrewind.sqlite (managed local DB path):

uv run main.py ingest-db --since-days 2 --browsers zen

Run the live vectorization health test (OpenAI-compatible /v1/embeddings on llama-server):

uv run tests/vectorization_concurrency.py

privacy

Privacy is a core goal of this repo: browser data stays local, processing is local-first, and the project is built to avoid shipping personal browsing data to third-party services.

Nothing here should require internet connections or API calls.