SiftStack

Full-stack real estate investing operations platform built around DataSift.ai.

Two things live in this repo:

1. The REI skill library. 24 Claude skills that teach Claude to run a specific REI workflow: comp a property, estimate a rehab, find the heirs, score your phone numbers, grade your cold calls. Install them into Claude Code with one command, or upload them to a Claude Co-Work session. You do not need the platform to use these.
2. The platform. The pipeline that pulls county distress data, enriches it through 10 steps, and pushes it into DataSift ready for sequential marketing. Requires a DataSift.ai account.

Start with the skills. Most people never need anything else.

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/DataSift-Ty-Personal/SiftStack/main/install.py | python3 -

# Windows PowerShell
irm https://raw.githubusercontent.com/DataSift-Ty-Personal/SiftStack/main/install.py | python -

That downloads all 22 current skills (24 packages counting the 2 plugins) into ~/.claude/skills/ and the 2 plugins into ~/.claude/plugins/. Restart Claude Code and they are live. No clone, no pip install, no virtualenv. Python 3.9 or newer, standard library only.

Prefer to read the script before running it? That is the right instinct:

curl -fsSL -o install.py https://raw.githubusercontent.com/DataSift-Ty-Personal/SiftStack/main/install.py
less install.py
python3 install.py

python3 install.py --list                                  # the catalog, with descriptions
python3 install.py --only rehab-estimator comp-package     # just these
python3 install.py --category "Deal Analysis"              # a whole category
python3 install.py --dest .claude/skills                   # into a project, not your home dir
python3 install.py --dry-run                               # report, write nothing

Re-running is safe. Anything already current is skipped, and a package is only swapped in after it downloads and unpacks cleanly, so a failed update cannot leave you with a half-written skill.

Co-Work takes the packaged file rather than a folder. Download the .skill you want from dist/ and upload it to your session or Project:

https://raw.githubusercontent.com/DataSift-Ty-Personal/SiftStack/main/dist/rehab-estimator.skill

Every package in the catalog below is at that same path with its own name. .plugin files work the same way.

If you are already in a Claude Code session, this is usually the fastest route:

Read https://raw.githubusercontent.com/DataSift-Ty-Personal/SiftStack/main/skills/manifest.json and install the Deal Analysis skills for me.

The manifest is the machine-readable index of the whole library: every package, its category, its description, its download URL, and a SHA-256 of the exact bytes. It is what install.py reads, and it is regenerated by CI on every push so it cannot drift from what is actually published.

24 current packages. python3 install.py --list prints this with full descriptions.

Two superseded packages (deep-prospecting, deep-prospecting-v4) stay in the repo so old runs still resolve. The installer skips them unless you ask for one by name. Use deep-prospecting-v5, which is about 5x cheaper and returns relatives on records where v4 returned nothing.

Eleven of the 24 work the moment they are installed. No key, no login, no card. Ask the library where you stand:

python3 install.py --doctor

It reads your environment and a .env if you have one, then prints what works now, what needs you signed in somewhere, and what needs a credential. Every blocked skill prints its no-API alternative on the same line, so you are never told no without being told what to do instead. It sends no requests, spends nothing, and never prints a credential value.

Every paid step has a no-API route. Comping runs by browser, heir research runs off obituaries and free people-search, KPIs run off a CSV export, coaching runs off any dialer's recordings. The no-API playbook has the method for each, and is honest about the three things that have no substitute.

Skills degrade rather than fail: a missing key means that step is skipped and the run says so.

The platform is 76 agents across 9 divisions.

• The interactive map, searchable and filterable by division, status, and whether a human still signs off.
• docs/AGENT-MAP.md, the same thing as a document you can read in the repo or hand to Claude.

Both render from docs/agents.json, so they cannot disagree about what the system does.

Each agent carries a trigger, the steps it runs, where a person signs off, what comes out, and the trap it exists to avoid. That last field is the one worth reading: every one is a real production failure, and most of them failed silently for days or weeks before anyone noticed.

  Web scrape (gated site)   ──┐
  Scanned PDF (OCR)         ──┤
  Courthouse photos         ──┼──►  Enrichment  ──►  DataSift  ──►  Sequential
  Dropbox auto-poll         ──┤     (10 steps)       (API)          marketing
  CSV re-import             ──┘
                                          │
                                          ▼
                          Deal analysis ──► Dispo ──► Lender package

Divisions: Data Acquisition, Enrichment and Identity, Deal Analysis, Dispo and Buyers, Outreach and Marketing, CRM Operations, Market Intelligence, Coaching and Performance.

Every gated action routes to a human. Outreach identity is anchored to the assigned person rather than a company name, and no agent sends a dollar figure to a seller or a buyer without a person approving it.

Requires a DataSift.ai account. Works with any market, any county, any state.

git clone https://github.com/DataSift-Ty-Personal/SiftStack.git
cd SiftStack
pip install -r requirements.txt
playwright install chromium
cp .env.example .env          # then fill in your keys

python src/main.py daily                          # new notices since last run
python src/main.py daily --upload-datasift        # scrape, enrich, upload
python src/main.py daily --counties Knox --types foreclosure,probate

All five produce the same records and flow through the same enrichment pipeline.

Deduplicate, filter vacant land, flag entity owners, resolve probate property, tax delinquency, address standardization (Smarty), property data (Zillow), obituary and heir research, skip trace (Tracerfy), phone scoring (Trestle).

Every API is optional. A missing key skips that step and the run reports it.

The county pull runs unattended on Fly.io rather than on a workstation. deploy/FTM_RUNBOOK.md is the runbook.

python src/ftm_runner.py --doctor        # credentials, egress, state, saved searches
python src/ftm_runner.py --max-notices 2 # bounded dry run, writes nothing
python src/ftm_runner.py --commit        # the real thing
python src/ftm_schedule.py --next        # next 5 fire times

The scrape is gated on egress IP, not on code. From an office IP the notice page carries no CAPTCHA at all, just a refusal to serve. From a datacenter IP it serves the normal challenge. A blocked run exits 3 rather than 1, because no retry fixes it and treating it as a normal failure sends you debugging the parser. Run --doctor before anything else.

Zero notices is a failure, not a quiet day. There is deliberately no inference that a missing challenge means a cleared gate. That exact reasoning reported 13 consecutive dead runs as successful over 19 days.

# Comps, ARV, rehab, buyers, into one workbook
python src/comp_package.py --address "158 Old State Rd" --zip 37914 \
    --beds 2 --baths 1 --sqft 1946 --bbox "35.996,36.016,-83.895,-83.840"

# The hour after you walk a house: 9 sheets, anchored to the live CRM record
python src/post_walkthrough.py --walkthrough-template
python src/post_walkthrough.py --address "..." --walkthrough walk.json

# The 8-piece private lender package
python src/lender_package.py --spec deals/EXAMPLE_lender_spec.json
python src/lender_docs.py    --spec deals/EXAMPLE_lender_spec.json

deals/EXAMPLE_lender_spec.json is a scrubbed, working spec. Copy it and fill it in.

Every variable is documented in .env.example. The ones that matter most:

Running one county daily costs roughly $40 per month in API spend.

The architecture is market-agnostic. Knox and Blount County, TN are what it is proven on.

1. Edit SAVED_SEARCHES in src/config.py for your county's notice site.
2. Add your county's tax API alongside Knox in src/tax_enricher.py.
3. The parser handles 7 notice types already: foreclosure, tax sale, tax delinquent, probate, eviction, code violation, divorce.
4. Photo import and the enrichment pipeline are county-independent and work nationwide.

The source of truth is the folder, not the ZIP. skills/<name>/ is what you edit and what gets reviewed; dist/<name>.skill is a build product.

# edit skills/rehab-estimator/SKILL.md, then
python tools/build_skills.py --build --manifest
python tools/build_skills.py --verify        # CI runs this

CI blocks a merge if a distributed archive drifts from its source, if the manifest is stale, if any SKILL.md is missing its name or description frontmatter, or if anything that looks like a credential appears in a package.

That frontmatter check exists because a skill with no frontmatter is silently undiscoverable. Claude has nothing to match a request against, so it simply never fires and no error is ever raised. One skill shipped that way for months.

1. Fork, branch, commit.
2. Run python tools/build_skills.py --build --manifest before you push.
3. Open a pull request.

MIT. See LICENSE.

Built by DataSift.ai for the REI community.