Sales Partner

Fresh, researched leads every week, each matched to the offer most likely to land, with sourced talking points ready for you to approve.

Most small businesses know they should be doing outreach. Few have the hours to find the right businesses, look into each one, and work out what to say. Sales Partner does that groundwork for you: it finds businesses that fit your ideal customer, checks them against your own criteria, researches the promising ones, picks the one offer from your own catalog that suits each, and writes talking points your own sequences and calls can use. You stay in charge of who gets contacted and when.

What you get each week

• New leads that fit. Businesses that match your ideal customer, found from map listings, websites and web search, each given a fit score against your own criteria with the reasons written down.
• Real research on the best ones. What's changing at each business (new hires, a new location, a missing website, recent news), who the decision-maker is, and how to reach them. Every fact links to where it was found. A second score, from that research, is recorded for you to read; it doesn't trigger anything.
• Evidence from reviews, ads, jobs and the website. It reads a business's recent Google reviews (quoting a snippet with its date, link and the reviewer's first name, never their profile or photo), checks whether it runs Google or Facebook ads, looks for open jobs on Indeed, and reads its website, guided by the research tips you gave for each offer. Each source is capped, and a lead's Apify spend stops at $0.50 by default. A lead that hit the cap carries a Research Capped flag and says which sources it skipped, so you can decide whether to continue.
• The offer most likely to land. Each lead gets a Segment, an Industry, and a Hero Offer from your offer catalog, with an Offer Reason and an Offer Evidence rating (strong, moderate or weak) that say how well the research backs the pick. When your rules rule out every offer, the Hero Offer stays empty and says why.
• Talking points to work from. For each researched lead it writes a Sales Partner · Talking Points note on the company in your CRM: who it is for, the offer, an opener, sourced points, what to avoid, and the ways to reach them. Your own sequences, automations or calls use them. Each run adds a dated section and never rewrites an earlier one.
• Social profiles for the people it finds. Confident matches are filled on the person record; possible matches go in a note on the person. LinkedIn is searched by name. Instagram, Facebook and TikTok come from the business's own pages, and from Exa when it is connected. Nothing is ever followed or messaged.
• A point of contact on each new lead. The most senior person it found with an email address is set on the lead, with the reason, and the talking points are written for them. Your email workflow (for example n8n) reads it. Set or change it by hand and it is never overwritten.
• A weekly summary sent to you: what's waiting for your review, what's approved but not yet contacted, what moved, and what research tools cost.

How it works

1. A short interview. It learns your business, your ideal customer, the offers you sell and the signals that point to each one, the results you can point to, and how you like to sound.
2. It works on a schedule. Each week it finds, scores and researches new leads, matches each to an offer and writes its talking points, all inside your CRM. It stops at Researched.
3. You review. Read the research, the offer and the talking points, and drag the leads you like to Approved. If a lead below the research bar deserves a look, tick Research Anyway and it is researched on the next run.
4. Your process takes over. From there, your own CRM automations, sequences, email tool, or you yourself make contact. Sales Partner's job is done, and it never touches a lead at Approved or later.

Why you can trust it

• It never sends anything to a prospect. It has no ability to email, message or post to them on your behalf. Nothing goes further than Researched until you approve it, and contact stays with you and the tools you already use. It messages only you, through Resend, from a no-reply address on your domain.
• You can see everything it did. Every run writes a record to an Airtable base you own. The agent can add to that record but never change or delete it.
• It never makes things up. Facts about a prospect come with a source link or are marked unverified. Claims about your business come only from what you told it.
• It respects "no." A business marked Do Not Contact is never prepared again, and the agent can't remove that mark.
• It's safe to leave running. Every system it uses has written rules it cannot break, checked on every action rather than just promised in its instructions. If a rule can't be checked, it won't run on a schedule.
• LinkedIn stays human on every platform. Talking points are text for you to use. The agent never follows, connects with or messages anyone on LinkedIn, Instagram, Facebook, TikTok or anywhere else.
• Your data stays yours. Leads live in your CRM, your business notes live in your own private folder, and passwords and keys never go into the agent.

What you need

• A CRM: Attio or HubSpot (or bring your own).
• Resend, your own account, for the digest and alerts the agent sends you (the free plan is enough), with a sending domain you verify.
• Airtable, your own account, for the agent's record of every run. Production needs the Team plan; Free (1,000 records) is fine for a trial.
• Claude Code on a paid Claude plan, which runs the agent and its weekly schedule.
• Apify, your own account, with a monthly usage limit you set in Apify Console, for local-business listings, website contact details, reviews, ads, jobs, website reading and social profiles.

Want it set up for you? Webspenser offers a one-time, done-for-you setup: we connect your tools, run the interview with you, and turn on the schedule. Ongoing management is available as part of our AI managed services.

---

Install (plugin mode)

/plugin marketplace add webspenser/agent-library
/plugin install sales-partner@webspenser

Then open Claude Code in an empty folder — this becomes your instance, where your data lives — and run /sales-partner:setup. It writes instance.yaml, runs the interview to fill context/, and offers to make the folder a private git repo. From then on, opening Claude Code in that folder loads the agent. Plugin updates never touch your folder.

Built on Agent Standard 7.5 with Agent Builder. Version 8.7.1.

Use it from source

1. Click Use this template to create your own private copy.

2. Clone it and run ./install.sh.

3. Run the setup skill in source mode. It writes instance.yaml at the root; the guard needs the agent line:

   agent: sales-partner
   mode: source
   

4. Ask the agent to run its interview (the interview-business skill).

Tools it needs

• A CRM — Attio or HubSpot (capabilities/crm/tools/); setup binds it after a read-only check. If a CRM needs fields the probe does not find, setup offers two choices: create them yourself from the tool's ## Setup steps, or run its bootstrap.py with an API key in your own terminal.

• Resend for owner notifications, on the client's own account (capabilities/owner_notify/tools/resend/); setup pins your address and a no-reply sender, and the guard refuses a send to anyone else.

• Airtable for the ledger (capabilities/ledger/tools/airtable/); bootstrap.py creates the base, and the guard keeps the record append-only.

• Apify for scraping (capabilities/scraping/tools/apify/): setup connects the restricted connector URL and pins the limits (results per search, pages per site, and so on) the guard enforces. The guard also requires the fixed settings, such as reviews with reviewer names and LinkedIn email search (your choice in 8.6.0), and checks every start URL: Facebook pages only on facebook.com, and only an https site root on a public host for the contact scraper and the site reader. One limit remains: the guard can't tie a website to a lead, so data could still leave in the hostname itself (a made-up subdomain). More actors find and read social profiles (LinkedIn by name, a LinkedIn company search and profile reader, Instagram and TikTok). Web search and page fetch read only https pages on public hosts with no query string; data could still ride in a page's path, the same class of limit as the hostname. They run with no login: the agent never signs in to any platform. Facebook personal profiles have no reader, so they come only from the business's own pages and from Exa. The guard limits which arguments go to the profile actors and how many, but it can't check what a name or handle contains, so data could in principle ride in one.

• Optional: Exa for web search (capabilities/web_search/tools/exa/), through Exa's hosted connector on your own Exa account. When it is connected, research runs one Exa search per lead to find its owner or principal; without it, research runs as before. The guard lets only Exa's basic search run, at its default 10 results.

Each tool's guard policy makes Attio, HubSpot, Resend, Airtable, Apify and Exa unattended-safe.

Connect these in your host (connectors or MCP servers). Credentials never go in this repo.

Scheduled runs: after setup, run /sales-partner:schedule.

Upgrading from 7.0: re-run setup's tools step to bind the ledger, commit and push bindings/ledger.md, then re-run /sales-partner:schedule. Until then, scheduled activities stop with an error notification.

Upgrading from 7.1: re-run setup's tools step to bind scraping, commit and push bindings/scraping.md, then re-run /sales-partner:schedule.

Upgrading from 7.2: the single Score is now two scores. The Fit score comes from the map listing and decides which leads get researched (research_threshold); the Approach score comes from research (since 8.0 it is recorded for you to read and gates nothing).

1. Rewrite the Scoring rubric in your context/icp.md as the two tables the shipped context/icp.md shows (Fit score, Approach score), or re-run the interview, and check research_threshold in context/operating-config.md against the new Fit rubric.
2. Re-run setup's tools step. The CRM probe now needs four fields (Fit Score, Fit Breakdown, Approach Score, Approach Breakdown); create them from the tool's ## Setup steps or re-run its bootstrap.py.
3. Score the leads already in flight once, in an interactive session: ask the agent to score every Scored lead with the Fit rubric (it re-reads each lead's listing with find_businesses, which counts toward the Apify spend cap) and every Researched lead with the Approach rubric (from its Research rows, adding the newest review date where it is missing). Until then, no gate picks those leads up.
4. The old Score and Score Breakdown fields are no longer read or written. They stay in your CRM as history. In Attio you can archive them (archived attributes keep their values). In HubSpot, archived properties are deleted for good after 90 days, so export the values first or leave the properties in place.
5. Commit and push the instance.

Upgrading from 7.3: research and drafts move into notes on the company, and every channel a lead can be reached on gets a full draft.

• Attio. Only the Sales Partner Pipeline list is used. Re-run setup's tools step (the probe now checks only that list). Then, before the next scheduled run, ask the agent in an interactive session to move research and drafts into notes: for each company, a Sales Partner · Research note with one section per original date, and a Sales Partner · Email, · LinkedIn or · Call note from the newest draft per channel. A pipeline phone or email that belongs to a linked person goes onto that person; otherwise it is recorded as a listing finding in the research note, with its source, so a lead whose only route was that phone keeps it. Then archive the Sales Partner Research and Sales Partner Outreach lists and the pipeline's Phone, Email, Address and Location attributes (archived data stays recoverable).
• HubSpot. New research and drafts are Notes on the company; existing draft Tasks stay as history.
• Commit and push the instance.

Upgrading from 7.4: Sales Partner no longer drafts first touches. It stops at Researched, matching each lead to an offer from your catalog and writing talking points. In order:

1. Update the plugin. Re-run the interview's offer catalog round and Round 2 (the customer and both rubrics), or the whole interview, to create context/offer-catalog.md and bring context/icp.md in line with it; the catalog round alone leaves icp.md as it was. Then check the instance's icp.md against the package's: Firmographics keeps no industry list, Industry fit scores against the catalog's industries, and Thresholds names only research_threshold (drop the approach_threshold bullet). Delete approach_threshold, enabled_channels, sending_identity and callback_phone from the instance's context/operating-config.md; nothing reads them now.
2. Re-run setup's tools step, so the bootstrap adds the new fields (Segment, Hero Offer, Offer Reason, Offer Evidence, Research Anyway), the catalog options and the Approved stage.
3. In the CRM, move leads at Ready to Send to Approved, and leads at Approach Drafted back to Researched. Then archive those two old stages.
4. Remove schedule_approach from schedules.yaml, then re-run /sales-partner:schedule. Until then, the schedule check refuses the stale entry.
5. For leads already at Researched with no Hero Offer, ask Sales Partner in an interactive session to match offers for them. That runs the re-match path, which first fills each lead's empty Segment from its Industry, or from its research. Old Email, LinkedIn and Call draft notes are left in place.

Upgrading from 8.0: Sales Partner now also reads reviews, ads, jobs and the website, and caps each lead's Apify spend. In order:

1. Optional: re-run the interview's offer catalog round to add research tips to each offer (the signals that tell the research where to look).

2. Replace the Apify connector's URL and re-run setup's tools step. The connector's restricted URL must now list all seven actors, so replace it with:

   https://mcp.apify.com?tools=compass/crawler-google-places,vdrmota/contact-info-scraper,compass/Google-Maps-Reviews-Scraper,scrapesage/google-ads-transparency-scraper,apify/facebook-ads-scraper,valig/indeed-jobs-scraper,apify/website-content-crawler

   The connector may now appear as "Apify Actors", with its tools prefixed mcp__claude_ai_Apify_Actors__; the guard still matches it. Then re-run setup's tools step. It adds the research_capped field (Research Capped), asks for the twelve new scraping bindings and pushes bindings/scraping.md.

3. Update context/operating-config.md: add apify_reviews, apify_ads, apify_jobs and apify_site_reader to prospecting_sources, add apify_cap_usd_per_lead: 0.50, and set research_budget_per_lead_minutes: 15.

4. Re-run /sales-partner:schedule.

Your scheduled routines install the latest plugin, so they run this release before you finish. Research goes on: a missing apify_cap_usd_per_lead uses 0.50. Until the tools step pushes the new bindings, the new sources are skipped and you get an error notification naming the missing bindings. Re-run /sales-partner:schedule afterwards (step 4): it checks the new bindings.

Upgrading from 8.1: decision-maker research can now use Exa, and every name it records is graded by its source (a directory-only name is never verified). Exa is optional. In order:

1. Without Exa: change the cloud environment's version comment to # sales-partner 8.2.0 and set agent_version: 8.2.0 in schedules.yaml, then commit and push. Nothing else changes. (8.2.1: use 8.2.1 in both places. 8.2.0's guard refused the objective argument Exa's connector requires, which stopped prepare at its first Exa search.)
2. With Exa: first add Exa's connector at claude.ai (the tool's ## Setup), re-run setup's tools step and answer yes for web search, then re-run /sales-partner:schedule, attach the ExaSearch connector to the routine that runs prepare, and do step 1.

Upgrading from 8.2: Sales Partner now finds social profiles for the people it researches. Exa is optional; with Exa bound, Instagram, Facebook and TikTok get more candidates. In order:

1. Replace the Apify connector's restricted URL with the ten-actor URL. Copy it exactly:

   https://mcp.apify.com?tools=compass/crawler-google-places,vdrmota/contact-info-scraper,compass/Google-Maps-Reviews-Scraper,scrapesage/google-ads-transparency-scraper,apify/facebook-ads-scraper,valig/indeed-jobs-scraper,apify/website-content-crawler,harvestapi/linkedin-profile-search-by-name,apify/instagram-profile-scraper,clockworks/tiktok-profile-scraper

2. Re-run setup's tools step. It creates the new CRM fields (with bootstrap.py or by hand) and pins the three new scraping bindings max_profile_results (5), max_profile_reads (10) and max_profile_locations (1). Then commit and push bindings/scraping.md. Until the new keys are pushed, scheduled runs stop.

3. Add apify_profiles to prospecting_sources in context/operating-config.md.

4. Re-run /sales-partner:schedule.

Upgrading from 8.3: each new lead now gets a point of contact on the pipeline list. In order:

1. Re-run setup's tools step. It creates the three new lead fields point_of_contact, point_of_contact_pick (the agent's own last pick; leave it alone) and point_of_contact_reason, with bootstrap.py or by hand.
2. Point your email workflow (for example n8n) at point_of_contact: in Attio it links to the Person, whose name and email you read.
3. Existing leads are not backfilled. Set theirs by hand.

Upgrading from 8.4: the digest now arrives as a formatted email, with the plain text still inside it, and no longer has a Next actions section. Add business_name: "<your business>" to context/operating-config.md so the title reads Weekly Digest for <your business>. Nothing else changes: Resend sends the HTML through the same pinned sender and recipients.

Upgrading from 8.5.0: the reviews source now runs on web_wanderer/google-reviews-scraper. Apify moved the old Compass reviews tool to an HTTP server interface the connector could not call, so every prepare run stopped at its first review call. Your prospecting_sources and bindings/scraping.md stay as they are.

1. Replace the Apify connector's URL with:

   https://mcp.apify.com?tools=compass/crawler-google-places,vdrmota/contact-info-scraper,web_wanderer/google-reviews-scraper,scrapesage/google-ads-transparency-scraper,apify/facebook-ads-scraper,valig/indeed-jobs-scraper,apify/website-content-crawler,harvestapi/linkedin-profile-search-by-name,apify/instagram-profile-scraper,clockworks/tiktok-profile-scraper

   Other named actors and the docs and runs tools may stay on the connector: the guard refuses any tool it does not allow. Leave off the actors and api tool groups. The connector's name must contain "apify" (for example "Apify Actors" or "Apify Connector"), because the guard finds the server by that word.

2. Start a new session so the agent loads the new tool.

Upgrading from 8.6.0: 8.7.0 adds nine Apify tools under the same guard: web search and page fetch (apify_web: press, news and pages from a lead's own site and search results, paths only), BBB (apify_bbb: rating, complaints, years, owners), YellowPages (apify_yellowpages: a listing for research, and leads by trade and city for the Prospector), Yelp reviews (apify_yelp), LinkedIn jobs (part of apify_jobs), a LinkedIn company search and profile reader (part of apify_profiles: the owner of a business that names no one, and their email) and a Facebook Ad Library backup (part of apify_ads, used only when the primary Facebook call fails).

1. Replace the Apify connector's URL with:

   https://mcp.apify.com?tools=compass/crawler-google-places,vdrmota/contact-info-scraper,web_wanderer/google-reviews-scraper,scrapesage/google-ads-transparency-scraper,apify/facebook-ads-scraper,valig/indeed-jobs-scraper,apify/website-content-crawler,harvestapi/linkedin-profile-search-by-name,apify/instagram-profile-scraper,clockworks/tiktok-profile-scraper,apify/web-search,apify/web-fetch,scrapesage/bbb-scraper,delicious_zebu/yellowpages-usa-business-lead-scraper,web_wanderer/yelp-reviews-scraper,valig/linkedin-jobs-scraper,harvestapi/linkedin-profile-search,harvestapi/linkedin-profile-scraper,scrapesage/facebook-ad-library-scraper

2. Run setup's tools step to add max_web_results (default 10), max_bbb_results (default 5) and max_web_fetches (default 5) to bindings/scraping.md, then commit and push it. Do this before the next scheduled run: the scheduled-run gate refuses to start while a value the guard pins is missing.

3. Add apify_web, apify_bbb, apify_yellowpages and apify_yelp to prospecting_sources in context/operating-config.md if you want them; a source left out never runs.

4. Start a new session so the agent loads the new guard.

Upgrading from 8.5.1: the Apify guard now follows the retrieval profile (Agent Standard 7.5). It names the actors it allows, caps every result and page count, refuses code, headers, cookies, proxies and each actor's paid add-ons, and lets any other new vendor input pass, so a small vendor change no longer blocks every call. Three behaviors change:

• A scraping source that fails (an empty field, a vendor error or a guard block) is skipped for that lead and research goes on; the run report says what happened. A source blocked on 2 leads is skipped for the rest of the run, and a source that fails on every lead sends one error notification at the end.
• Review findings and talking points can name a reviewer by first name and date. Profile links, photos and IDs are never kept.
• The LinkedIn search runs in Full + email search mode, and an email it finds for an owner, a Gmail address included, can become the point of contact. Its cost estimate is $0.15 per search; check the real charge in Apify Console after the first runs.

Your connector URL, prospecting_sources and bindings/scraping.md stay as they are. Confirm that the Apify monthly usage limit is set in Apify Console: it is the hard stop on spend. Start a new session so the agent loads the new guard.

Developing

tests/run-all.sh                                   # content checks
/path/to/agent-builder/bin/validate-agent.sh .     # structure

CI runs on pushes to main and on every pull request. Content tests and the release rule run only in webspenser/sales-partner — in your own copy, CI checks structure only, so committing your filled-in context/ is fine.

License

Apache-2.0.