teft-leads 0.1.0

Local-first LinkedIn intent-signal lead-finder with a local web app — real signals, not stored databases.

teft

teft (Norwegian: the instinctive nose for a lead — a scent for who's in-market right now). Repo slug is still linkedin-lead-finder; the tool is named teft.

A personal, local-first tool that turns a freeform brief into a ranked, explained shortlist of high-intent LinkedIn prospects — "real signals, not stored databases."

For colleagues — try it in 2 minutes

Nothing to clone. It runs entirely on your own machine (the web app is 127.0.0.1-only).

1. Install uv: curl -LsSf https://astral.sh/uv/install.sh | sh (Windows: see docs.astral.sh/uv)
2. Run: uvx teft-leads serve — your browser opens the app. (PyPI name is teft-leads; the command you type once installed is still teft.)
3. In the app: Settings → paste your keys → Search → run it.

That's it. The rest of this page is for people building on teft.

About / planning

Planned with the wayfinder workflow; the plan lives as GitHub Issues. Two maps:

• Planning map (#1) — complete. Destination reached: the locked v0 definition.
• Build plan (#6) — stack, LLM, architecture, name, milestones.
• Implementation map (#7) — active. v0 as five wayfinder:task milestones (M0→M4); the frontier (open, unblocked, unassigned) is the work takeable now — currently M0 (#8).

See CONCEPT.md for the original seed brief.

v0 at a glance

Python CLI on your Mac, over your own logged-in LinkedIn session (Patchright, li_at cookie reuse).

Guided: teft run — one command walks you through config (prompts + saves any missing key/cookie), then product → brief → buyer-intent phrases (or reuse the last campaign), scrapes, ranks, and opens the results UI. Defaults to headless/fast.

Or step by step:

• teft scrape -p "phrase" … — content-search → per-candidate activity/profile → candidates.json.
• teft rank --brief brief.txt — each candidate → the "why now" prompt, SKIP filter → ranked shortlist-<date>.{md,json}.
• teft view — a local, zero-dependency results UI: lead cards, expandable post, Open profile / Open post / copy email, and contacted + notes saved to ~/.teft/actions.json.

No Fit %, no monitoring, no outreach, no external services. .env is auto-loaded (no manual source). Speed/stealth knobs: TEFT_HEADLESS=1, TEFT_MIN_DELAY / TEFT_MAX_DELAY, TEFT_MAX_CANDIDATES (cap per query, default 25).

LLM provider

rank is provider-pluggable. Default is claude-opus-4-8 (the tier the "why now" prompt's SKIP-rule calibration was validated on). To use GPT-5.6 instead, set TEFT_LLM=openai (defaults to gpt-5.6-terra; override with TEFT_MODEL, e.g. gpt-5.6-sol / gpt-5.6-luna) and OPENAI_API_KEY. The SKIP calibration is only proven on Claude — re-run the fixtures on a new provider before trusting it (see VALIDATE.md).