Let Claude/Codex run actual 1000s of LinkedIn outreach with ~30 MCP tools

I'll pull the gist and HN thread so the paragraphs stick to stated facts only.A GitHub Gist from ayush488-glitch published Outreach Desk, a reusable SKILL.md that turns Claude or Codex into an operator for real LinkedIn outreach rather than advice-only chat. The HN post frames the setup as running actual thousands of LinkedIn outreach actions through about 30 MCP tools. The skill wires those agents to Second Brain Labs (app.secondbrainlabs.com, mcp.sbl.so) so campaign state can be read and changed from the chat client. On Hacker News the thread sits at 2 points with 3 comments.

Technically the agent is the reasoner and the execution surface is a remote MCP server exposing exactly 30 unique sbl_* tools over Streamable HTTP at https://mcp.sbl.so/mcp. The documented loop is define audience, build draft, source leads, analyze, review, approve, launch, then work replies. Lead sources include post engagement (preview free, import billable; preview capped at 25 people), Sales Navigator URL import, CSV up to 20 MiB, and a prompt-led ICP session where create/refine/get are free and approve is paid. Campaign build uses sbl_create_campaign_from_prompt for a DRAFT with opener and timed follow-ups plus an AI quality score and gap list. Launch and send are gated: sbl_run_campaign and sbl_send_campaign_message require explicit user approval; the skill defaults to stopping before launch and expects a literal LAUNCH token with frozen sender, recipient count, messages, campaign ID, and revision. Bulk copy uses template placeholders; real personalization is reserved for approved one-to-one replies.

For engineers and builders this is a concrete pattern for agent product design: put irreversible and billable steps behind tool-level approval gates, ground readiness claims in live reads (sbl_get_campaign, sbl_list_campaign_users) instead of chat memory, and separate free draft/read tools from paid import and launch tools. Installation paths cover Claude Desktop (DXT), Claude Code, Codex desktop and CLI, plus other MCP clients with an Authorization Bearer header. Known MCP limits are called out in the skill: LinkedIn message text cannot be patched in place over MCP, regenerating after import orphans leads and can re-bill, and there is no remove-lead tool so dirty list cleanup is dashboard-only.

In market terms this sits at the intersection of coding agents (Claude, Codex) and sales-execution SaaS, competing with manual Sales Navigator workflows, CSV-based sequencers, and dashboard-only outreach tools by moving the operator loop into the same MCP-enabled client developers already use for code. The honest scale note in the skill matters: many campaigns across segments are possible, but each billable import and each launch stays separately approval-gated, and LinkedIn or account limits still apply. Post-engagement is framed as small high-intent lists run often, not one giant scrape, which undercuts a naive reading of the HN title as unbounded automation.

Practical takeaway: treat the skill as an approval-gated ops playbook, not a fire-and-forget bulk sender. Connect MCP, confirm exactly 30 sbl_* tools, settle message copy at create-time before import, bind a LinkedIn sender with sbl_bind_linkedin_channel, and only launch after a live readiness checklist. Watch whether message component edits and lead attach/remove gaps close on the MCP surface, how post-engagement vs Sales Nav vs prompt-led sourcing behave in production costs, and whether the human-intervention queue plus sbl_reply_and_resolve becomes the real daily driver once campaigns are live.