Newsbrew

A private, local-first news aggregator built with SolidStart, Solid, TypeScript, shadcn-solid-style components, and TanStack AI. It uses an OpenAI-compatible local model server such as LM Studio.

What it does

1. Polls your RSS and Atom sources.
2. Sends each new headline, byline, and source through one stateful tri-state filter session. YES is clearly wanted, MAYBE is neutral or ambiguous, and NO is an explicit rejection. Newsbrew reads the model's active and maximum context lengths from LM Studio, tracks exact Responses token usage after every turn, and starts a fresh filter chain before the next turn would exceed the active window.
3. Fetches and analyses both YES and MAYBE articles. MAYBE stories remain chronologically inline with a more compact presentation; NO stories are suppressed.
4. Analyses each accepted article in a two-turn stateful session: first the headline, quick summary, and tags; then detailed Markdown points.
5. Combines private natural-language guidance with positive and negative topic signals. You can rate each story topic independently and feed those signals back into future filtering prompts.

Everything is stored locally under ~/.config/wes-dev/newsbrew/. The runtime source of truth is news.sqlite, accessed directly through Node's built-in node:sqlite module.

Requirements

• Node 24 or newer
• An OpenAI Responses-compatible model endpoint, such as LM Studio

Run with npx

Once the package is published, start Newsbrew without a repository checkout or an installed dependency tree:

npx newsbrew

The package ships the complete Nitro application and its runtime dependencies inside .output/. By default it listens on port 3000; use PORT and HOST to change the listener:

HOST=127.0.0.1 PORT=4321 npx newsbrew

From this repository, npx . exercises the same executable after the package has been built. npx newsbrew --help prints the runtime options without starting the app.

Runtime, model, source, topic, and authentication settings are stored in SQLite. Access is open by default; an optional shared access token can be set in the Settings drawer or imported with the rest of the configuration.

Settings can be imported from the default ~/.config/wes-dev/newsbrew/newsbrew.json, from a file selected by NEWSBREW_CONFIG_FILE, or directly from the JSON stored in NEWSBREW_CONFIG_JSON. Inline JSON takes precedence over a configured file. Imports are transactional and only run again when the JSON content changes. Relative databaseFile values in a settings file resolve from that file's directory.

NEWSBREW_CONFIG_FILE=/run/secrets/newsbrew.json npx newsbrew

Or provide the same object directly:

NEWSBREW_CONFIG_JSON='{"runtime":{"pollIntervalMinutes":30,"maxItemsPerSource":8}}' npx newsbrew

Omitting databaseFile uses ~/.config/wes-dev/newsbrew/news.sqlite. A leading ~/ is also expanded in configured paths.

Development and settings commands

Development requires pnpm. To initialise a database or force a configuration snapshot back into an existing database, pass its path to the import command:

pnpm settings:import -- ./newsbrew.dev.json

Export the current database-backed settings to the default ~/.config/wes-dev/newsbrew/newsbrew.json, or pass another destination:

pnpm settings:export
pnpm settings:export -- ./newsbrew.dev.json

Exports are written atomically with owner-only file permissions. Since access tokens are stored as one-way hashes, export preserves one only when the destination already contains its plaintext value; otherwise auth is omitted instead of silently disabling an existing token.

The command validates the JSON, opens the database selected by databaseFile, transactionally applies the settings, and prints a non-sensitive import summary. newsbrew.dev.json is ignored by git.

See newsbrew.example.json for the complete structure. Set auth.accessToken to a shared token or leave it as an empty string to disable authentication. Newsbrew stores a salted token hash in SQLite and accepts the token through the login screen or an Authorization: Bearer header. Configuration files and the SQLite database should be treated as sensitive because they may contain API keys or an access token.

Commands

pnpm dev
pnpm ingest
pnpm tune:filter
pnpm tune:analyser
pnpm benchmark
pnpm worker
pnpm check
pnpm test
pnpm build

• pnpm dev runs the SolidStart app.
• pnpm ingest performs one feed scan.
• pnpm tune:filter fetches and judges every candidate from the enabled feeds through one stateful filter session, logging every YES, NO, or MAYBE result as JSONL. It is read-only: it does not add articles, mark stories as seen, or change topic preferences. Use -- --limit=2 to reduce the number per source, or -- --source=ars-technica to test one source. Redirect stdout if you want to keep a clean JSONL log, for example pnpm --silent tune:filter > filter-run.jsonl.
• pnpm tune:analyser fetches full articles and runs the two-turn analyser, logging the exact article input, each structured model response, response ID, token count when provided by LM Studio, and timing as JSONL. It is also read-only. The same -- --limit=2 and -- --source=ars-technica options apply. To save a clean log, use pnpm --silent tune:analyser > analyser-run.jsonl.
• pnpm benchmark compares downloaded local models with the production tri-state prompt and stateful session shape. It reads the private reader guidance, topic profile, and labeled reference set from ignored ~/.config/wes-dev/newsbrew/benchmark-* JSON files and writes an atomic private report there. See BENCHMARKING.md before preparing or interpreting a run.
• pnpm worker scans immediately and then uses the polling interval stored in the database.

Both ingestion commands run TypeScript directly with Node's --experimental-strip-types flag. The TypeScript configuration enables erasableSyntaxOnly, and runtime imports use explicit .ts extensions so the worker stays compatible with native Node type stripping.

pnpm build creates the self-contained .output/ application. pnpm pack runs that build before producing the npm tarball; the published package has no runtime dependency installation because Nitro has bundled those dependencies into the output.

See ARCHITECTURE.md for the durable pipeline and session decisions, and BENCHMARKING.md before comparing models or prompts.

Notes

• Add local publications by pasting their RSS or Atom URL into the Sources panel.
• Publisher-provided article images are preferred. When one is unavailable, the app uses a non-generated, topic-related image at half height and labels it as related.
• Some publisher pages block automated reading. Those stories are skipped and the worker continues.
• Keep the model server bound to localhost unless you intentionally secure it.
• Respect publishers' terms and use this for personal reading.