← All Automations MARKETING INTELLIGENCE CLI

Brand Pulse

A single-file CLI (with an optional terminal-styled web UI) that gives marketers a real, honestly-scoped brand snapshot across Reddit, Twitter/X, YouTube, and web/news — synthesized into themes with real, clickable citations, not just a raw mention count.

WHAT IT DOES

Noise in. Signal out.

• Fans out a live search for any brand name across Reddit, Twitter/X, YouTube, and web/news in one command — real scraped results, not a cached index
• Every count is a scoped result-set size ("top N results in the last M days"), printed next to the number — no fake precision search APIs can't actually back up
• A second Claude pass (Haiku 4.5) groups the sampled results into up to 3 themes and writes a one-line brief — no invented numbers, no hallucinated trends
• The model cites its supporting sources by item number only, never by writing a URL itself — every "real citation" claim is structurally guaranteed, not just asserted
• Built on agent-reach, which picks the best available access method per platform — no official paid APIs required

HOW IT WORKS

One command, five steps.

1. Step 1 — Multi-Platform Fan-Out

   For each platform — Reddit, Twitter/X, YouTube, web/news — a real search runs against the brand name and pulls back the top ~25 real titles and URLs. No API is queried once and reused; each platform gets its own live call.

2. Step 2 — One Numbered Pool

   Every title from every platform gets merged into a single numbered list, tagged with its source platform — the raw material the synthesis step will actually work from.

3. Step 3 — Constrained Synthesis

   The full numbered list is handed to Claude in one shot (via the local claude -p CLI, Haiku 4.5): discard anything not plausibly about the brand, group what's left into up to 3 themes, write a one-line brief, and cite up to 3 supporting items per theme — by item number only.

4. Step 4 — Citation Mapping

   Cited item numbers get mapped back locally to their real (platform, title, URL) tuples. The model never writes a URL itself, so a cited link is always one that was actually scraped in Step 1 — hallucinated citations aren't just unlikely, they're structurally impossible.

5. Step 5 — CLI or Web UI

   Run it as a one-line CLI command, or spin up the optional local FastAPI frontend — a terminal-styled page where each platform resolves live, and the finished brief renders with clickable citations.

SEE IT RUN

A real check, live.

A real Duolingo check, mid-synthesis — 4 platforms resolved, themes rendering with clickable, real citations.

UNDER THE HOOD

Open source, zero lock-in.

1. Single-file CLI + optional web UI

   brand_pulse.py runs standalone; app.py adds a local FastAPI frontend. Python 3, MIT licensed, public on GitHub.

2. No paid APIs required

   Reddit, Twitter/X, and YouTube access route through agent-reach's per-platform tooling (OpenCLI, twitter-cli). Web/news search runs through Exa via a bundled, keyless mcporter config.

3. Cost-efficient synthesis

   The synthesis call runs on Claude Haiku 4.5 through the local claude CLI — a repeatable, low-stakes summarization task that doesn't need a heavier model.