Streetlight

Free, no-account AI support for people navigating housing instability, paperwork, benefits, and hard decisions in Seattle / King County.

Anonymous by default. No accounts, no PII retention, no ads, no upsell, no data harvesting, no premium tier. Open source from day one.

A software engineer in SOMA uses Claude every day. He pays $20/month and it's a rounding error in his budget. A guy sleeping in a shelter ten blocks away has the exact same brain. Often harder problems. Zero access. That gap is not a moral failure on his part. It's an access gap. This project closes it, in one city, for one population, as a public utility.

Status

Pre-launch. V1 in development. Soft-launch through named frontline partners when the build is ready.

Documents

• Agent instructions — shared coding-agent entry point for Claude Code, Codex, and other agents.
• Thesis — why this tool exists and the reasoning behind it.
• V1 Specification — what the tool is, concretely.
• Data and Privacy Architecture — how it's built, what it logs (and doesn't), and the threat model.
• What this tool will not do — the "no" list.
• Translation handoff — how locale files are produced, reviewed, and audited for completeness.
• Translation worklist — generated per-language missing-key checklist for translation maintenance.
• Resource maintenance — how the hand-curated human-help lists are verified and kept fresh.
• Operational runbook — dashboard-first incident and pause procedures.
• Partner bug report template — structured no-content issue intake for frontline partners.
• Partner launch packet — ready-to-send framing, limits, and bug-report links for soft launch.
• Partner materials index — quick map of the soft-launch docs set.
• Security policy — how to report vulnerabilities safely.
• Contributing guide — how to open issues and send pull requests.
• Code of conduct — participation and moderation standards.

Project structure

This is a one-person, one-year commitment with a graceful sunset path if no institutional home is found by year-end. The about page is honest about this. Code stays public on GitHub regardless.

Useful commands

• npm run lint
• npm run dev:mock — runs the app with zero-cost local mock chat replies
• npm run dev:live — runs the app against the configured live model path
• npm run build
• npm run check:audit — runs npm audit for dependency advisories
• npm run check:locales
• npm run check:locales:summary — quieter locale-gap summary for routine verification loops
• npm run check:launch — static soft-launch readiness check for docs, placeholders, translations, and env contract
• npm run check:env — validates local/env-example boolean and numeric value shapes without printing secrets
• npm run check:forbidden-integrations — fails if forbidden analytics, error-reporting, or log-forwarding packages are added
• npm run check:no-tracked-secrets — scans tracked files for common secret-shaped values before push
• npm run check:ops — dry-run validation for the incident and ops helper scripts
• npm run check:response-style — runs live brief-first response-shape cases against the configured /api/chat path
• npm run check:translation-worklist — verifies that the generated translation worklist is current
• npm run cost:status — shows the local model pair and whether cheap local mode is active
• npm run diagnostics:local — prints the current local server mode and the report-page metadata snapshot
• npm run ops:status — prints the local health, deploy metadata, model mode, and any launch-doc placeholders still hanging around
• npm run build:translation-worklist
• npm run incident:new -- --slug short-name --severity Sev-2 — scaffolds a public Sev-1/Sev-2 incident write-up from incidents/TEMPLATE.md
• npm run incident:sev3 -- --what "..." --action "..." --outcome "..." — appends a structured Sev-3 note to incidents/log.md
• add --dry-run true to either incident command if you want a preview without writing files
• npm run validate:data — validates referral/crisis schema plus source and last-verified metadata
• npm run data:status — shows the oldest referral/crisis entries and whether any have drifted past the freshness threshold
• npm run smoke — checks key pages plus a small synthetic end-to-end chat flow against or

npm run dev:mock is for zero-cost local UI work. npm run dev:live is the easy way back to real model calls. Full prompt regression still expects live model replies, so npm run regression:prompts will stop with a clear message while mock mode is active. Use npm run regression:mock for a full no-spend plumbing pass. The GitHub Actions Verify workflow runs the full mock suite on PRs and pushes to main; run npm run regression:prompts with the Haiku testing model path before model, prompt, or deploy changes where live behavior matters. Run npm run check:response-style after prompt changes that affect answer length or structure.

Utility routes

• /healthz — simple no-store JSON health check for local smoke checks and future monitoring
• /report-problem — structured no-content bug report draft builder that stays client-side and opens a mailto: draft

License

Licensed under the Apache License, Version 2.0. See LICENSE.

Contact

Bug reports currently go to [email protected].