chezy

Finding a flat, made easy.
The AI rental concierge that reads listings like a local, spots the traps, and calls the agency for you.

You describe the flat you want in a chat. chezy searches real Barcelona listings, checks each one forensically (does "luminoso" mean sunlight or a 1.5 m lightwell? is the price above the barrio average?), and when you approve one it phones the agency in Spanish, discloses that it is an AI assistant, agrees a viewing slot, and drops the confirmed viewing on your calendar.

The listing works for the agency. chezy works for the renter.

Quick start

With your coding agent

Using Claude Code, Cursor, Devin or another agent? Paste this:

Clone https://github.com/espetro/chezy and set it up: run `mise trust`, copy
apps/web/.env.example to apps/web/.env.local and fill in OPENAI_COMPATIBLE_API_KEY with
my Nebius key (leave VIEWING_MODE and CALENDAR_MODE on "mock"), then run `mise run setup`
and `mise run dev`. Open http://localhost:3000, complete the onboarding, and confirm that
listing cards render for "a bright 2-bed in Gracia under 1800".

By hand

You need mise on your PATH and your keys in apps/web/.env.local (start from apps/web/.env.example). Everything else is pinned and installed for you.

mise trust          # install node, pnpm, python, uv, pg0 at the pinned versions
mise run setup      # pnpm install, uv sync, start pg0, migrate, seed 300 listings
mise run dev        # http://localhost:3000

Only OPENAI_COMPATIBLE_API_KEY is required for the chat. VIEWING_MODE=mock and CALENDAR_MODE=mock keep the call and the calendar deterministic; switch them to slng/vonage and google once you have those credentials.

What happens in a session

Beat	What you see
Brief	"Find me a bright 2-bed in Gracia under 1,800." Onboarding fills the missing pieces (areas, budget, bedrooms) conversationally.
Shortlist	Real Barcelona listings, ranked against your profile, rendered as cards inside the chat.
Forensic check	One card carries an amber flag: 0% direct sunlight, windows onto a lightwell, 13% over the barrio average. Another passes.
Approval	You like the passing one and tap "call the agency".
The call	A voice agent dials the agency, speaks Spanish, opens with the AI disclosure, and agrees a slot.
Booked	The slot lands on your Google Calendar and the chat confirms the time.

Rejections are remembered: tell it why a flat is out and the next round reranks.

Why

Two facts about renting in Barcelona compound each other:

• Listings lie. In Q1 2026 asking rents on idealista averaged about 20 EUR/m² while registered contracts averaged 16.89 EUR/m² (INCASOL via Spanish Property Insight).
• Speed decides. 36% of Barcelona flats rented through idealista in Q2 2026 were gone in under 24 hours, versus 17% nationally (idealista/data).

If you are new in town, do not know barrio price norms, cannot read a lightwell from the photos, and cannot call agencies in Spanish during your workday, every step of the funnel is stacked against you. chezy is the agent on your side of the table.

Message chezy on Telegram

Bot handle: @hackbarna_chezybot. Open the link on your phone, say hi, and it onboards you in chat (areas, budget, bedrooms), then searches Barcelona rentals and offers to call the agency for a viewing (approval card first). Built on Mastra (@mastra/core, model router over OPENAI_COMPATIBLE_* env, Telegram via Mastra Channels) for the HackBarna 2026 Mastra challenge.

apps/bot (@chezy/bot) is a second client for the same concierge: a Mastra agent reachable from Telegram, running alongside apps/web and reusing its listings and Postgres store. It is an adapter, not a fork: the bot imports apps/web/lib read-only and, since PR #75, adapts the web tools (searchListings, saveUserProfile, recordListingFeedback, approval-gated arrangeViewing) instead of keeping bot-local ones. It adds only Telegram-specific pieces (identity mapping, approval cards, radar alerts) in a bot Postgres schema.

Every turn on web and Telegram lands in the audit log as a chat.turn.* record: mise run audit:turns tabulates .audit/ JSONL (channel, model, latency, tools, cited listings).

How it maps to the "build an agent people can message" criteria:

• Works from a stranger's phone: polling mode, no webhook or tunnel needed. Any Telegram user can DM the bot; TELEGRAM_ALLOWED_USER_IDS stays empty.
• Memory: Mastra Memory on @mastra/pg (dedicated mastra schema), 20 last messages plus a working-memory template that persists budget, neighbourhoods, must-haves and red lines across sessions.
• Approval-gated actions: arrangeViewing never runs on its own. The agent proposes, Telegram shows an approve/deny inline keyboard, and only approval dispatches the call (mocked with VIEWING_MODE=mock).
• Proactive: a radar loop re-scores buildFeed for every linked Telegram user and DMs fresh matches above RADAR_MIN_SCORE, deduped via bot.radar_seen. mise run bot:radar triggers a pass on demand.

Run it locally:

cp apps/bot/.env.example apps/bot/.env   # add TELEGRAM_BOT_API_KEY + TELEGRAM_BOT_NAME
mise run db:start
mise run bot:dev

Headless verification uses apps/bot/test/fake-telegram.ts, an in-memory Bot API server: mise run bot:test covers a plain DM, the approval-gated viewing flow and radar dedupe with no network.

Built with

              

One file per sponsor challenge lives in docs/evidence/.

Layer	Choice	Evidence
App	Next.js 16, React 19, AI SDK 7, shadcn/ui, Tailwind 4	evidence index
LLM	Nebius AI Studio through @ai-sdk/openai-compatible (DeepSeek-V4.1-Flash by default)	nebius.md: providers, models, embeddings, vision
Voice	SLNG managed agent on LiveKit SIP, Vonage Voice API for the PSTN leg	slng.md: 2 live calls + transcripts + usage; vonage.md: live PSTN receipt
Bot	Mastra agent on Telegram (@chezy/bot)	mastra.md: live @hackbarna_chezybot
Data	Drizzle + postgres-js on pg0 (embedded Postgres 18 + pgvector), 300 committed Barcelona listings	evidence index
Calendar	Google Calendar via service account	evidence index
Scraper	uv-managed Python CLI (httpx, parsel, pydantic) with fotocasa, habitaclia, idealista, milanuncios and pisos.com adapters	pisos.com adapter written by Devin, see below
Autonomy	Chezy Forge: Devin sessions driven through the v3 API, verified by pytest + hidden hold-outs, failures fed back until the verifier passes (scripts/devin-forge)	Cognition (Devin) track evidence: write-up, raw logs, PRs #66 #70 #71
Evals	Galtea and QualityClouds runs against the concierge	galtea.md: find/fix/prove deltas; norma-qualityclouds.md + DEFENSE.md: 62 → 67/100
Observability	LogTape + JSONL audit, one chat.turn.* record per turn on web and Telegram	architecture, mastra.md

Cognition (Devin) track evidence

Chezy Forge creates Devin cloud sessions through the Devin v3 API, verifies what Devin ships with pytest, pydantic, ruff, basedpyright, hidden hold-out fixtures and a path allowlist, and feeds failures back into the same session until the verifier passes. No person or model decides.

• Write-up with both run timelines, the run logs pasted verbatim and the judging-criteria table: .agents/notes/2026-09-20-devin-forge-run.md
• Plan: .agents/plans/2026-09-20-devin-adapter-forge.md
• The layer: scripts/devin-forge/ (forge.ts loop, devin.ts v3 client, verify.ts gates, tasks.ts standards, prompt.ts), run with mise run forge:pisos or mise run forge:pisos-detail
• Standards Devin had to meet and the hold-outs it never saw (now regression tests): apps/scraper/tests/test_pisos.py, test_pisos_holdout.py, test_pisos_detail.py, test_pisos_detail_holdout.py, test_pisos_detail_holdout2.py, fixtures under apps/scraper/tests/fixtures/ (pisos_*)
• The artifact Devin built: apps/scraper/src/chezy_scraper/adapters/pisos.py, a fifth listing source (uv run scraper scrape --platform pisos --operation rent --tier small pulled 61 live listings)
• Raw machine logs: .agents/evidence/devin-forge/
  • run-20260920-1239-list-adapter.jsonl.txt: session created via API, verdict pass, merged
  • run-20260920-1304-detail-parser.jsonl.txt: created, pass, merged, session resumed, attempt 0 FAIL fed back (seed), attempt 1 pass, merged
  • gate-logs/20260920-1304-attempt-0-pytest.log: the real first failure (4 failed, ValidationError: heating), gate-logs/20260920-1304-attempt-1-pytest.log: 17 passed after Devin's fix; plus uv_sync, ruff_check, ruff_format, basedpyright logs per attempt
  • forge-stdout-run1.log, forge-stdout-run2.log: the forge process output
• Pull requests Devin opened and the forge verified and merged: #66 search-page adapter, #70 detail-page parser, #71 the retry that fixed a crash found on a live page; umbrella #73
• Devin sessions (origin api): run 1, run 2 with the fed-back regression

Repo map

apps/web         the product: chat, onboarding, explore, viewing + calendar routes
apps/bot         Telegram client (Mastra + polling adapter), second surface for the concierge
apps/scraper     listings pipeline and the committed dataset builder
apps/video       Remotion demo video
packages/*       contract (Valibot), db (Drizzle), config, ui, observability
chezy-mock-data  the 300-listing Barcelona snapshot (JSONL + Parquet) plus derived enrichment tables
assets/          brand: logo at every size, Outfit and DM Sans fonts

Docs

• PRD.md: problem, demo narrative, goals, scope, hackathon tracks
• ARCHITECTURE.md: pointer to the module graph and boundary rules
• AGENTS.md: enforced gates and conventions for humans and agents
• assets/README.md: brand assets and how to load them
• apps/web/README.md, apps/scraper/README.md, apps/video/README.md

Status

Hackathon build, September 2026. The closed loop (brief to booked viewing) runs on the committed dataset with mock voice and calendar; the live SLNG call and the evaluation runs are in flight. Not production software: Barcelona only, single user, no real auth.

Submission cuts

Two annotated tags mark what to judge:

• latest — current branch HEAD; moves forward as merged, green work lands. Refresh with git tag -fa latest -m "<one-liner>" HEAD from the merged commit.
• deadline — the last commit submitted before 14:00 CEST on the hackathon submission day. Immutable judging reference; never moved after the cutoff.

Inspect either with git show <tag> --no-patch --format="%H %ai %s" to see the exact commit, author timestamp, and subject.