1 Signed-in session
/chrome-debug-profile + Playwright on CDP — your SSO, cookies, apps. The foundation that worked across orgs.
you type · AI runs the rest · grounded
Full AI workflow.
Local cloud included.
LocaDev
is the loop where the agent does almost everything: gather requirements,
clarify them in Slack / Discord / Teams, update
Jira and Azure DevOps boards,
open PRs with gh, and exercise
desk-hosted Azure & AWS–shaped resources — while
pre/post hooks refuse decisions and “ready” without citations and receipts.
You type. You approve. The machine ships the busywork.
# clone the GitHub repo (source of truth)
git clone https://github.com/gtfodevs/locadev.git
cd locadev
# ground → decide → build → prove ready
./hooks/pre-decision.sh
make up && make verify
./hooks/post-ready.sh
git clone https://github.com/gtfodevs/locadev.git
cd locadev
# ground → decide → build → prove ready
./hooks/pre-decision.sh
make up && make verify
./hooks/post-ready.sh
The whole party (AI-driven)
End-to-end automation except the human at the keyboard.
Browser-first: your signed-in Chrome session (no API keys for every SaaS) —
then site skills, grounding, local cloud, and gh.
2 Gather anywhere
Not only Jira fields: product UIs, chat threads, PDF/Excel in Teams/Slack, wikis. /web-requirements plans, then snapshots.
3 Boards in the browser
Read/comment/transition Jira & Azure DevOps in the UI you already use. API CLI optional. Layer thin site skills for repeat flows.
4 Ship with gh
Assumes GitHub CLI is installed and authed: branches, PRs, checks — agent-driven, human-approved.
5 Local cloud buffet
Summon only the Azure/AWS-shaped bits you need: blobs, bus, SQL, Redis, Functions-style workers, S3, mail, channel fakes — env-only wiring.
6 Grounded “ready”
/grounding demands snapshot citations before decisions and receipts before ready — not vibes.
Jira + Azure DevOps
Default: browser UI via CDP (no tokens). Optional REST CLI when headless. Details: boards/README.md · pattern: browser-skills.md.
Browser
Open Jira/ADO in your Chrome session. Read AC, comments, attachments. Comment & transition in the UI. Snapshot for citations.
Jira / ADO API
Optional: ./boards/board.sh with JIRA_API_TOKEN / AZURE_DEVOPS_EXT_PAT when you want the CLI path.
Site skills
Repeatable org flows as thin skills on top of chrome-debug + Playwright — not a new OAuth app per tool.
Pre / post hooks (no hopium)
Big choices and “we’re done” both fail closed without grounding. Protocol lives in hooks/ on GitHub.
⚓ Pre-decision
Before architecture, board rewrites, or committing work: cite docs, channel messages, jira:/ado: keys, or repo paths. ./hooks/pre-decision.sh
✓ Post-ready
Before “ready for review”: verify, tests, Jira/ADO update, gh PR evidence. ./hooks/post-ready.sh
📎 Citations = truth
Empty sources fail the gate. Conflict? Ask the channel again — don’t assume and ship.
What the desk runs
Same SDK shapes as production. Env-only costume change when you finally go real-cloud.
☁ Azure cosplay
Azurite, Service Bus, Foundry-shaped bridge, optional Cosmos & AI Search, Key Vault, sample Functions on Azurite.
⬡ AWS cosplay
MiniStack S3 on :4566. endpoint_url, test keys, no surprise invoice.
💬 Channel cosplay
Fake Slack / Discord / Teams so agents can post clarifications and you can see the messages land.
✅ Receipts, not vibes
Playground, demos, captured mail, message UIs — so “it works” survives a demo. UI changes: /playwright e2e for post-ready.
Browser-first (the proven stack)
Stack chrome-debug + Playwright, then build site skills on top — drive tools you already use (boards, chat, PDFs) without API keys. Full map: docs/browser-skills.md · local.
🔐 chrome-debug-profile
Base session. Work-copy of your Chrome + CDP :9222. SSO, cookies, Jira, Teams — the login problem solved once.
🧪 playwright (session)
Drive that session. Click boards, open chat PDFs/Excel, capture pages. Clean Chromium is only for your app’s CI e2e later.
📋 web-requirements + site skills
Gather & specialize. AI plans, then snapshots under .grok/local/requirements/. Org skills layer on the same CDP base.
Pick your power-ups
Core stays light. Profiles are optional loot — equip only what this run needs.
| Profile | What you get | Port |
|---|---|---|
core | Azurite, Service Bus, bridge, PGlite, Topaz, Redis | — |
aws | MiniStack S3 gateway | 4566 |
search | Qdrant + AI Search emulator | 8800 |
cosmos | Cosmos DB vNext emulator | 8081 |
mail | Fake SendGrid (+ /captured) | 8095 |
slack | Fake Slack Web API (+ /ui) — clarify in-channel | 8096 |
discord | Fake Discord REST (+ /ui) — clarify in-channel | 8097 |
functions | Azure Functions runtime + sample (Azurite storage) | 7071 |
teams | Fake Teams + echo bot — clarify in-channel | 3979 |
sample | In-repo FastAPI consumer | 18080 |
Start the machine
Everything ships from github.com/gtfodevs/locadev — clone first, then summon the stack. macOS / Linux + Docker Compose v2.
1 · Clone from GitHub
git clone https://github.com/gtfodevs/locadev.git · star/watch the repo for updates · open issues there.
2 · Tool up + summon
Agent host + gh auth + boards config. Then make up && make verify — wire from sandbox.env.example.
3 · Loop with hooks
Browser gather (chat/PDF/tickets) → cite → code → ready. You type; AI drives the session.
Not magic — just honest
Local cloud is a sidekick, not a free Azure/AWS subscription in a trench coat. Topaz ≠ Entra. PGlite ≠ multi-writer Postgres. AI Search is partial. Hooks force citations so the AI doesn’t paper over those gaps — or invent requirements. Source of truth: github.com/gtfodevs/locadev.