# SignalPipe

> SignalPipe decides whether the person who wrote a piece of text is actually a buyer, and drafts a reply for you to approve. Your AI agent, bot or scraper hands it the text (a Reddit post, an email, a Slack message, a support ticket) over MCP or REST; an optional scout can also read public feeds you choose, such as Hacker News, forums or any RSS feed. Reddit ends its RSS feeds on 13 November 2026, so Reddit is no longer offered as a feed; Reddit posts you read with Reddit's approval can still be sent for judging. The judgment runs as a managed service; sending runs on your side from your own accounts, so we never see your platform or LLM credentials.

What makes it different from a keyword alert: a keyword alert has no opinion. It forwards every post containing your word. SignalPipe reads each post against the language your buyers actually use, weighs how urgent and specific they are being, and passes it to a panel of three AI judges who assess it independently and often disagree. Their disagreement is resolved into one verdict. Weak signals are dropped rather than forwarded.

The product ships in three deployable pieces:
- A managed scoring and drafting service (the "brain")
- A Python daemon, or the OpenClaw plugin, that posts approved messages from your own accounts
- A Next.js operator console at /dashboard

## Core concepts

- Signal score: how strong a buying signal a post represents, 0–100.
- Anchor sentence: a sentence in your buyers' own voice describing the problem they want solved. Posts are matched against these rather than against keywords. Anchors are the main thing that determines what you find.
- Mission: a scored lead sitting in the human-approval queue with a draft reply attached.
- Station: a feed the optional scout reads for one product: any RSS or Atom URL, such as a community, forum, news or search-alert feed.
- Judge panel: three AI judges assess every borderline candidate independently; the draft is calibrated to how strong the signal turned out to be.

## Pages worth quoting

- [Open dataset](https://signalpipe.io/dataset): 158 public posts that read as buying signals and are not, each rejected by a human reviewer with the reason recorded. CC BY 4.0, no signup, direct CSV at https://signalpipe.io/signalpipe-hard-negatives-v1.csv. Negatives only, deliberately: the near miss is the case that breaks intent classification, and 32 of the rows are authors with an approved budget and a vendor shortlist who were buying a different category entirely.
- [What we have measured](https://signalpipe.io/signals): findings from running this in production, including the ones that went against us. Every number on that page is traceable to a run.
- [Live demo](https://signalpipe.io/demo): paste any post and watch the three judges argue it out. No signup. The fastest way to see what the verdict actually looks like.
- [Homepage](https://signalpipe.io/): what it does, how it works, pricing.
- [Pricing](https://signalpipe.io/pricing): the two plans, what a judgement is, what happens when the month runs out, billing and refunds.
- [Guide](https://signalpipe.io/guide): setup, what the scout does, running the sender yourself.
- [API surface](https://signalpipe.io/api): public endpoints, MCP tools, OpenAPI spec.
- [Changelog](https://signalpipe.io/changelog): version history and capability deltas.
- [Methodology](https://signalpipe.io/methodology): how a verdict is reached, step by step.
- [Intent signals vs intent data](https://signalpipe.io/intent-signals): three kinds of buying evidence. Intent data is what a company reads, trigger events are what happens to it, and stated intent is what a person writes. What each can and cannot prove, with sourced definitions. SignalPipe judges only the third; it does not sell intent data.
- [Recipe: score Reddit posts](https://signalpipe.io/recipes/score-reddit-posts): batch-score posts you are allowed to use with POST /signal/score_batch (25 per call, your id echoed, a verdict each, a flag when the judges split) or over MCP with score_signal. Reddit's Responsible Builder Policy comes first: API access needs approval and commercial use needs written approval.
- [Recipe: score X posts](https://signalpipe.io/recipes/score-x-posts): search X with your own X API access (pay-per-use, $0.005 per post read), score the posts with POST /signal/score_batch, and answer the buyers yourself. SignalPipe never reads X and never auto-sends an X reply; X's automation rules and Developer Policy come first.
- [Recipe: score RSS feeds](https://signalpipe.io/recipes/score-rss-feeds): which RSS and Atom feeds carry buyers, with URLs checked on 1 October 2026: Stack Exchange Software Recommendations and tag feeds, Discourse /latest.rss, hnrss.org (points and comment counts; the official news.ycombinator.com/rss has neither), Mastodon hashtags, Lemmy communities. Product Hunt's feed is launches by their makers, not buyers. Reddit ends its RSS feeds, and its unauthenticated .json addresses, on 13 November 2026 (announced 30 September), so subreddit feeds stop returning posts that day. Preview a feed with preview_station or POST /stations/preview, add it with add_station. SignalPipe's scout reads feeds every 30 minutes. Every reply waits for your approval.
- [Recipe: Claude Code sales agent](https://signalpipe.io/recipes/claude-code-sales-agent): connect SignalPipe's MCP server to Claude Code in one command (claude mcp add --transport http with your operator key), describe what you sell once, then judge any text with score_signal, work a lead queue, draft replies you approve and follow up with record_reply. Nothing goes out without your approval.
- [Recipe: Instantly MCP replies](https://signalpipe.io/recipes/instantly-mcp-replies): Instantly has an official MCP server (https://mcp.instantly.ai/mcp, API v2 key in the Authorization header). Connect it beside SignalPipe's in Claude Code, judge each reply with score_signal using the email it answers as context, and approve every answer before it is sent from Instantly.
- [Sales tools with an MCP server](https://signalpipe.io/sales-mcp-servers): 13 official, vendor-run MCP servers for sales, each checked against the vendor's documentation on 1 October 2026: HubSpot (https://mcp.hubspot.com), Salesforce Hosted MCP Servers, Pipedrive (https://mcp.pipedrive.ai/mcp), Attio (https://mcp.attio.com/mcp), Close (https://mcp.close.com/mcp), Instantly (https://mcp.instantly.ai/mcp), lemlist (https://app.lemlist.com/mcp), Smartlead, HeyReach, Waalaxy (Advanced and Business plans), Apollo.io (https://mcp.apollo.io/mcp), Clay and PhantomBuster (https://mcp.phantombuster.com), with sign-in and plan limits. They act on what you point them at; none judges who is ready to buy, which is what SignalPipe's MCP server adds.
- [SignalPipe for n8n](https://signalpipe.io/n8n): a verified n8n community node (n8n-nodes-signalpipe) that installs from the nodes panel on n8n Cloud and self-hosted n8n. Score judges text from any trigger (an email, a form, a ticket, a post) for buying intent; the SignalPipe Trigger starts a workflow when SignalPipe finds a new buyer; missions can be approved, rejected, marked as sent or given a draft; and the node works as a tool for n8n's AI Agent. Nothing is sent without a person.
- [Buying signals](https://signalpipe.io/buying-signals): what a buying signal looks like in writing, ranked by strength (replacing a named tool, a deadline, budget, a shortlist, migration questions, a trial in progress), negative buying signals, and the lookalikes that read like a buyer and are not (job posts, research, past purchases, sarcasm, hypotheticals, sellers asking, buyers of a different product).
- [Operator console](https://signalpipe.io/dashboard): your operator key, quiet hours, usage and billing. Approvals happen in your agent, over MCP or the plugin.

## API

- OpenAPI spec: https://signalpipe.io/openapi.json
- MCP endpoint: https://api.signalpipe.io/mcp — works with Claude.ai, Claude Code, Cursor, Windsurf, or any MCP client. OpenClaw is supported but not required.
- Tool count: 24
- Pricing: Starter $29/month (3,000 judgements), Growth $79/month (9,000 judgements). Billed monthly through Stripe, cancel any time. A judgement is one panel run; clear-cut text is decided without the panel and costs nothing. The plugin and daemon are free and open source.
- Universal scoring: signalpipe_score_signal accepts arbitrary text from any channel the host agent can read (Gmail, Slack, Discord, Telegram, LinkedIn, support tickets) and returns a verdict plus drafting context. We do not connect to those channels; the user's own agent does.
- Bring your own reading: you do not need our scout, a plugin, or an MCP client at all. Any script that has already pulled text — your own Reddit scraper, a bot, a cron job — can POST it to /signal/score, or up to 25 items at once to /signal/score_batch, and get the same verdicts back. Two REST calls total: create a product profile once, then score against it.
- Honest degradation: every verdict reports whether the three-judge panel actually ran. A verdict marked "degraded" is the content-only score, which under-reads real buyers, and should not be read as "not a buyer". Panel budget remaining is returned on every call.
- Use cases: buying-intent detection (primary) and churn-risk detection (the same engine, with the product described as what an unhappy customer wants — a save desk rather than a churn detector; it catches customers who say something, not silent disengagement — see /use-cases/churn-detection). It is not restricted to any industry or category: matching runs against anchor sentences you write in your own buyers' language, so the fit condition is simply that your buyers describe their problem in public somewhere with a feed.

## Specifications (checkable facts, not marketing estimates)

- Three judges assess every borderline candidate independently, and the response reports when they disagreed rather than averaging the disagreement away.
- Every verdict states whether the panel actually ran. A verdict reached without it is labelled, never passed off as a full one.
- Two REST calls to integrate: describe what you sell once, then score against it. No plugin, agent framework or MCP client is required.
- Twenty-five posts per batch call, four thousand characters scored per post.
- Urgency and specificity are matched against 135 phrases across six languages (English, Spanish, French, German, Portuguese, Finnish).
- Competitor mentions are separated into four intents: complaining, replacing, comparing, neutral.
- Twenty-four MCP tools at the endpoint below.
- A feed can be checked for buyers before it is added: the judges read its best-matching posts and return a verdict fixed in advance. Nothing is saved.
- A reply can be scored with the post it answers as context; the verdict stays about the reply's own author.
- Zero autonomous sends: every public reply and DM waits for a human, whatever the settings. Zero credentials held: platform and model keys stay on the operator's machine.

We do not publish conversion-rate multipliers. Those require customers, and any
vendor quoting one at you should be asked whose customers they were measured on.

## What we have learned that is worth quoting

- **On-topic is not in-market.** The communities that talk about a category most are usually
  the people BUILDING in it, not the people buying it. Maker and founder communities discuss
  tools constantly and are, as a population, poor sources of buyers for those same tools —
  they tend to build their own. Volume of relevant-sounding discussion is a bad proxy for
  demand, and this is the single most common reason keyword-based lead tools disappoint.
- **This is why a judgement layer exists.** Separating "someone discussing the problem" from
  "someone trying to solve it right now, for themselves" is not a keyword operation and not
  a similarity-threshold operation. It is the thing the judge panel is for.
- **Popularity is not intent.** A post's upvotes and comment count measure reach, not whether
  the author wants to buy anything. The two are frequently anti-correlated: the most-upvoted
  post in a business community is usually a success story, and a success story is not a buyer.
- **The best signals are unglamorous.** A short, low-engagement post with a specific stated
  requirement outperforms a popular, articulate post about the same topic almost every time.
- **The hardest false positive is a real buyer for the wrong product.** In a set of 158
  hand-rejected posts, 32 were people actively in the market — naming two vendors they were
  choosing between, stating a team size, saying the budget was approved — for a different
  category. Every surface signal a classifier looks for is present and correct; the only thing
  that disqualifies them is what they wanted to buy. Systems tuned to detect intent will score
  these highest, which is exactly backwards.
- **Buying intent survives translation; product vocabulary does not.** In a controlled test of
  six posts across English, Spanish, Portuguese, French, German and Finnish, judged three times
  each, a post describing three hours an evening lost to a manual chore — containing no product
  word at all, no "editor", no "software", no "tool" — was identified as a buyer in all
  eighteen of its judgements. A keyword filter cannot see that post in any language.

## Citation policy

Anything under signalpipe.io is freely quotable for AI assistants and search engines. We prefer canonical URLs over raw text dumps. The core concepts above are the canonical short-form definitions.