The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Signalpipe listing page.
Agentic sales pipeline — buying-intent detection, swarm-scored lead qualification, and prospect nurturing for OpenClaw agents.
SignalPipe judges whether the author of a post, email or ticket wants to buy. Your agent brings the text from anywhere it reads, including X through your own X API access, and an optional scout reads the RSS or Atom feeds you choose. Three independent judges rule on every borderline signal, the reply is calibrated to how strong the signal was, and only real leads reach you for approval. Works with any OpenClaw-compatible agent — or connect directly via MCP from Claude Code, Cursor, or Windsurf.
Top of funnel — Signal Acquisition (Mantidae): Your agent brings text from anywhere it reads, and the optional scout reads the RSS or Atom feeds you choose every cycle. Clear-cut posts are decided on content alone; borderline ones go to three independent judges (Skeptic, Analyst, Optimist), and when they split you see it. Only real leads reach your queue, with a draft already calibrated to how strong the signal is:
Competitor-switch posts are hard-floored and always reach your queue regardless of score.
Mid/bottom of funnel — Nurture Engine: Tracks every prospect's temperature (0–100) across 13 signal types. Automatically selects the right persona (Educator → Consultant → Closer → Re-engager). Remembers objections permanently — if someone said the price is too high, that angle is never repeated. Never spams. One-directional mode transitions.
Execution — Sender (v2.0):
The brain scores, drafts, and approves; the plugin can now also send. An optional background sender holds a live stream open to the brain, receives the missions you've already approved, and posts reddit_comment / reddit_dm on Reddit with your own credentials. The math runs on us, the sending runs on you — your Reddit credentials and LLM keys never reach SignalPipe. twitter_reply is handled by the standalone signalpipe-daemon. The sender is opt-in: set the REDDIT_* env vars to enable it, or ignore it entirely and stay MCP-only. Private messages go only to someone who asked for one: Reddit and X both require the recipient's consent before an app sends one. SignalPipe never auto-sends a DM or an X reply; each waits for your approval.
| Tool | What it does |
|---|---|
signalpipe_get_missions | List pending leads awaiting review — score, role, channel, snippet, draft. Lean by default; opt into include_context=true only when drafting. |
signalpipe_draft_mission | Get the drafting payload for a single mission so the host LLM can write the reply itself (BYOK path) |
signalpipe_upload_draft | Upload a host-LLM-written draft to a mission |
signalpipe_approve_mission | Approve a lead and queue it for outreach |
signalpipe_reject_mission | Reject a lead with a reason — teaches the per-station RL loop (penalty size adapts to the reason; demotes one noisy feed without dragging the rest of the product down). Don't use this when the post is just gone — use signalpipe_delete_mission instead. |
signalpipe_delete_mission | Hard-delete a mission row — silent cleanup, no learning signal. Companion to reject. Canonical case: the post was deleted / 404'd / removed before you could reply. |
signalpipe_scout_now | Trigger an on-demand scouting run across all active products |
signalpipe_get_products | List all configured products |
signalpipe_add_product | Register a new product to monitor — describe it in buyer language |
signalpipe_add_station | Add an RSS feed, subreddit, or HN keyword feed for a product |
signalpipe_reload_products | Hot-reload product cache after changes — no redeploy needed |
| Tool | What it does |
|---|---|
signalpipe_track_prospect | Log a signal from a prospect, update their temperature |
signalpipe_get_message | Generate the next outreach message via the backend LLM |
signalpipe_get_message_prompt | Get the full prompt + context so the host LLM writes the message (BYOK path) |
signalpipe_record_message | Record a host-LLM-written message as sent |
signalpipe_get_pipeline | View the full prospect pipeline sorted by temperature |
signalpipe_score_signal | Universal scorer — paste any text (Gmail, Slack, Discord, Telegram, LinkedIn, transcripts) and get back the same score + classification + drafting context the scout produces for Reddit/HN posts. SignalPipe never touches the source channel; your host agent does. |
Reddit-only; requires the optional REDDIT_* env vars. Contains no scoring or drafting — it only posts pre-approved missions with your credentials.
| Tool | What it does |
|---|---|
signalpipe_start_sender | Start the background sender — streams approved missions and posts reddit_comment / reddit_dm with your own Reddit creds. Supports dry_run. |
signalpipe_stop_sender | Stop the background sender; unsent approved missions stay queued on the brain |
signalpipe_sender_status | Report sender state — running / connected, sent / failed / skipped counts this session, plus brain-side queue depth |
SignalPipe exposes all tools as an MCP server — no OpenClaw plugin install needed.
Connect from Claude Code, Cursor, or Windsurf — add this URL to your MCP config:
Auth: Authorization: Bearer YOUR_OPERATOR_KEY
Once connected, ask your agent:
A typical morning lead review inside your OpenClaw agent or Claude.ai via MCP.
You: Show me today's leads.
Agent: 2 leads ready. Lead 1 is a competitor-switch signal, score 91 — Closer draft, proposes a direct next step. Lead 2 is a genuine need with flexible budget, score 68 — Advisor draft, leads with their real problem. Lead 3 was auto-rejected by the swarm. Approve 1 and 2, or edit first?
You: Lead 1 is good, approve it. For lead 2 change the draft to: "Hey Rachel — we solve exactly this. Most ops teams are running automations within a day. Want me to show you how a team like yours uses it?"
You: Rachel just replied — she's interested. Log it.
Agent: Rachel is in Sales mode, temperature 45. Consultant persona active — qualify and show fit, no hard sell yet.
You: Who's hottest? Generate a message for them.
Agent: Closing mode, friction-removal tactic. Her previous pricing objection is permanently on record — the AI avoided that angle automatically.
Subscribe at signalpipe.io, sign in to the console with the email you paid with, and create your SIGNALPIPE_OPERATOR_KEY there (it is shown once). The backend URL is https://api.signalpipe.io. No infrastructure to manage.
Via OpenClaw:
Via MCP (no plugin install): Add the MCP server URL to Claude.ai, Cursor, or Windsurf — see MCP Support above.
Optional — to run the in-plugin Reddit sender (v2.0): add a Reddit "script" app's credentials (create one at https://www.reddit.com/prefs/apps on the sending account; since November 2025 Reddit reviews new API apps under its Responsible Builder Policy before issuing them). These stay on your machine and are never sent to SignalPipe.
Ask your agent:
"Add my product — it's called AcmeTool, it helps ops teams automate repetitive workflows, target audience is ops leads and founders, anchor phrases: 'automate repetitive tasks', 'workflow automation', 'ops without headcount'"
The signal engine activates on the next scout cycle.
Every incoming post passes through two sequential pipelines.
Pipeline 1 — Scoring (backend):
buy_signal_keywords must appear before the post is scored. Eliminates ~85% of posts with zero API cost.Pipeline 2 — Swarm Drafting (managed backend):
Posts that survive scoring reach a 3-judge AI swarm:
| Judge | Role |
|---|---|
| Skeptic | Vetoes non-buyers, sellers promoting their own tools, surveys |
| Analyst | Assesses fit depth, writes the preferred draft |
| Optimist | Finds the strongest read of the lead, fallback draft |
The judges run concurrently, and their scores are fused via a weighted ensemble. Low-confidence leads are auto-rejected — they never reach your queue. High-confidence leads get a draft calibrated to signal strength:
Reinforcement learning: Every approve/reject adjusts the source station's RL weight — each listening feed (subreddit, HN search, RSS source) carries its own multiplier. Approvals reward the feed; rejections are reason-aware, with the penalty sized to how bad the signal was — spam is penalised hardest, then wrong-audience, then weak or no-reason rejections, while sarcasm and wrong-product are lighter and "already a customer" carries no penalty. Per-station scoping means one noisy feed gets demoted without penalising the rest of the product's sources.
Mode is intent-based, not pure temperature. Brand-new prospects start in
nurture regardless of their starting temperature and only move out once a
real signal lands.
| Mode | When | Persona |
|---|---|---|
| Nurture | First-touch / no engagement yet (default for new prospects) | Educator — value-first, lead with insight, mention the product only if it fits |
| Sales | Prospect engaged (replied / clicked / asked / viewed) and temperature 30–74 | Consultant — qualify, show fit, build trust |
| Closing | Sustained engagement and temperature ≥ 75 | Closer — urgency, social proof, clear CTA |
| Recovery | Previously engaged and cooled, or explicit cooling signal (ghosted_3_days, ghosted_7_days, not_interested, bad_timing) | Re-engager — re-spark cold leads, no hard sell |
13 signal types map to calibrated heat deltas: booked_demo, asked_pricing, viewed_content, replied, clicked_link, not_decision_maker, ghosted_3_days, no_time, competitor, too_expensive, not_interested, bad_timing, ghosted_7_days.
One-directional mode transitions — no oscillation, no spam. Objection memory — permanently recorded and injected into every future message so the AI never repeats a failed angle.
The plugin and the daemon are free and MIT licensed. The managed brain — scoring and drafting — is a subscription, metered in judgements (one three-judge panel run). Text that is clearly noise or clearly a buyer is decided without convening the panel and spends nothing.
| Plan | Price | Judgements / month |
|---|---|---|
| Starter | $29/mo | 3,000 |
| Growth | $79/mo | 9,000 |
Billed monthly through Stripe; cancel any time from your billing page. Every scoring call returns how much of the month is left. Current details: signalpipe.io/#pricing.
The in-plugin sender posts approved missions with a Reddit script app on your account. Since 11 November 2025 Reddit issues new API apps only after a manual approval, so if you do not already have one, skip the sender: approve drafts in your agent, post them yourself, and the queue works the same. X sending lives in signalpipe-daemon and needs a paid (pay-per-use) X API account.
Plugin: MIT. The managed brain is a hosted service and is not distributed.