v1.4.0: MCP v2, Apollo's agent, and three new skills

Apollo shipped a four-tool MCP router, opened the MCP to master API keys, and put its own AI agent behind nine MCP tools. This release tests all three on a live account and rewrites the lane model around what held. Three new skills bring the library to 20.

MCP v2: prefer the router

https://mcp.apollo.io/mcp-v2 exposes four tools instead of 74: 5.6 KB of schema instead of 344 KB, every v1 tool still dispatchable, identical search results. The hand-maintained tool map is now a convenience and apollo_find_tools is the authority. The dispatcher is a free cost signal: anything routed through apollo_write_destructive spends credits or changes live state, so it gets a cost quote and a human first.

A master API key runs the MCP unattended, with no consent prompt and no session expiry. It also reaches tools the router documents but refuses to dispatch (record collections, CSV exports, dynamic AI enrichment, domain diagnosis), which we reported to Apollo.

Apollo's own agent: complement, do not compete

Nine apollo_agent_* tools are nine doors to one agent. We tested it. Planning is free, and it is good: it read the Context Center, planned four campaign concepts, sized four audiences live, and stopped for confirmation.

It is strong at the front of the motion and absent from the back: no deliverability check, no independent verification, no composition grading, no suppression, no warmup clock. That is the half this library exists for. Where the two overlap on targeting, planning, or copy, the skills now say the option out loud and let you choose. Whatever the agent returns still goes through List and Infrastructure before anything sends. Product how-to questions route to Apollo's own knowledge, not ours.

Reachable today only on the v1 endpoint with a master key. Check apollo_find_tools rather than trusting the changelog.

Three new skills

• campaign-reporting: results per sequence, step, variant, mailbox, seniority, and day of week, plus meetings booked and opportunities. Read-only, verified against the Apollo UI.
• apollo-account-audit: 14 read-only checks on an existing account, graded, stops at findings. Its first run found an active sequence sending nothing because its touches were never approved.
• account-based-outbound: when to reach several people at one company, how many, and which account tools carry it. Tested live.

Also new

• A compliance reference: opt-outs, the workspace GDPR setting, do-not-call screening, recorded-call consent. Operating rules, not legal advice.
• Install as a plugin from any directory: /plugin marketplace add creatop-gtm/apollo-operator, then /plugin install apollo-operator@creatop. About 1,400 always-on tokens per session.
• Company search, measured: 1 credit per call including the empty page, and a company row already carries founded year, revenue, NAICS, SIC, and headcount growth over 6, 12, and 24 months, so growth and sector grade off one free join. Plus a zero-credit preview sample for a prospect or client.
• Two targeting traps with numbers: a founder title filter returns Founding Engineers and Designers too (9 to 14% of a raw list), and "recently funded" includes acquisitions (34% of one universe).

Corrected with dates

• Sequences are readable now. apollo_emailer_campaigns_show returns every step and body; v1.3 said no lane could.
• MCP can stop a live sequence, with a trap: the update is declarative, so any step you omit is deleted. The CLI's sequences abort stays the kill switch.
• Step 3 is the shortest email after all, one or two sentences. v1.3 had reversed it.
• Single contact create matches on email and overwrites the existing record. Bulk create never updates on a match. Never test with a real address.
• AI-written touches are silently dropped on create. Read every touch back.

Full detail, with dates and measurements, in the CHANGELOG.