An opinionated go-to-market operator for Apollo.io: open-source skills for Claude Code, Codex, and any AI chat.
Apollo's MCP puts your whole Apollo account inside your AI assistant: search, enrichment, sequences, and sending, all in plain language. Apollo's own plugin adds a strategy skill on top, and it is good. What no tool can hand you is the judgment an operator brings to every step: whether a list is worth sending, when a domain is ready, whether the copy reads like a person, and what the results actually mean.
Apollo Operator is that judgment, written down. It is 20 skills that sit beside Apollo's own tools and run outbound on the Apollo MCP in order: understanding the business, who to target, how to build and grade a list, how to write sequences that get replies, how to build the sending infrastructure and take campaigns live without torching your domain, and how to tell whether any of it is actually working.
One router skill (operator-context) loads first and hands off by phase. A phase is a kind of work, not a step you finish once.
Foundations (always on)
operator-context: the tool map, the rules that protect your sending reputation, and the routing.apollo-operator: the four ways to run Apollo headless and when to reach for each.
Context (business-brief): one readable brief.md any agent can read once and understand the business. Optionally projected into Apollo's own Context Center.
Targeting (apollo-icp-builder): turn an ICP into a real Apollo search, size several angles for free, and pick one on numbers instead of instinct.
Infrastructure (sending-infrastructure, apollo-deliverability): dedicated domains, mailboxes, DNS, warmup, the volume math, and the bounce rules. This phase sits early on purpose. Warmup is a 14 to 21 day wall-clock wait, so the clock has to start while you are still building the list. Teams that leave it until the copy is approved discover a finished campaign they cannot send for three weeks.
List (apollo-list-builder, list-quality-scorecard): search, grade composition, suppress, enrich, verify, and grade. Nothing reaches a sequence until it passes.
Message (apollo-sequence-builder, sequence-reviewer, and optionally apollo-multichannel): short sequences matched to the persona, reviewed for spam risk, plus LinkedIn, call, and one-off email steps.
Launch (apollo-go-live): enroll, a human approves, contacts pull on reply, and there is a kill switch for when something goes wrong.
Iterate (positive-reply-scoring, experiment-design, weekly-rhythm, re-engagement): measure intent, run clean experiments, keep a cadence, and go back to the 90% who never replied.
Signals (cross-cutting, apollo-signals): find the reason to reach out now.
Most people know Apollo has an MCP and an API. There are four routes to the same data, and picking the right one decides how much context you burn and whether a job is possible at all.
| Lane | Auth | Results go | Best for |
|---|---|---|---|
| MCP (prefer v2) | OAuth, or a master API key | Through context | Guardrailed, human-gated steps |
| CLI binary | apollo auth login |
To disk | Bulk work with common filters |
| REST + the CLI's own token | Reuses the CLI's token, no API key | To disk | Bulk work needing advanced filters |
| REST + API key | API key | Anywhere | Building a separate service |
The third is the one most people miss. apollo auth login stores an OAuth token, and that token authenticates against the same REST endpoint the MCP uses. So you get the complete MCP filter surface (NAICS, tenure, headcount growth, department headcounts) while writing results straight to disk. Measured against MCP on an identical query: same total, 3 MB to disk, nothing in the context window.
Neither surface contains the other, which is worth knowing before you commit to one. Mailbox purchasing and the Context Center are MCP-only. Deals and rate-limit stats are REST and CLI but not MCP. And Apollo's MCP v2 is a router: four tools instead of the full catalog (85 on 2026-09-25, and growing), about 1% of the context cost, and the same search results.
The rule it collapses to: search wide off-context, act narrow on MCP.
We build this on our own Apollo account, and running it is what finds the problems. Nothing here comes from reading the docs: every item below cost us something to learn. Some of what testing caught, now encoded:
- A title filter for
founderalso returns Founding Engineer, Founding Designer, and Founding GTM: 9 to 14% of a raw list, and none of them buyers. - "Recently funded" includes acquisitions. Apollo's latest funding event can be the company being bought: 34% of one funded universe was Merger / Acquisition. Always pair the date filter with the funding stage code.
- A data provider's catch-all flag is not a verification. On one list it reported clean on all 846 addresses while an independent verifier found 39% catch-all.
- MCP can stop a live sequence, and can wipe it doing so.
apollo_sequences_updatewithactive: falseworks, but the update is declarative, so any step you do not echo back is deleted. Use the CLI'sabortin an incident. - Running several angles against one ICP overlaps silently: 85 people in a second list had already been enriched for the first.
- A title filter quietly inflated a search from 679 real matches to over 91,000.
- A provider's own
verifiedstatus is not deliverability. 997 addresses gradedverifiedcame back from an independent verifier as 564 safe to send, 413 catch-all, and 3 that would have bounced. Verification is a separate control, and it belongs before you write copy, not after. - Bulk contact creation does not deduplicate, despite documenting that it does. The same five contacts submitted twice produced two records each. Never retry on a timeout.
- Enrollment can fail silently. Passing several contact ids to the CLI joins them into one string, so Apollo finds nothing and returns exit 0 with an empty campaign. Assert on the response body, never the exit code.
- Apollo's MCP router runs ahead of its dispatcher, and the dispatcher catches up. A tool appears in
apollo_find_toolsfirst, runs on the v1 MCP with a master API key next, and lands in v2's dispatch list and the docs last; the CLI's login cannot reach it until that last step. Eleven tools made that journey between 2026-09-14 and 2026-09-25, record collections and CSV exports among them. The nine agent tools have not yet. Record collections also need a plan gate Apollo currently calls "Sheets", which is not a plan you can buy today. - You pay per record attempted at enrichment, not per email found, and re-enriching a record you already paid for charges again.
- Credits are eleven separate pools, not one balance, and the AI pool is 200 times the size of the contact-data pool. Data is the scarce resource, not generation.
Outbound has a lot of jargon and a lot of ways to quietly waste money, so the operator is built to work with someone who has never done this before. It names what an artifact is missing and why, glosses jargon on first use, gives every number a baseline, and states what a step costs before running it.
If a step would spend more than 85% of your remaining credits, it stops and offers you the choice rather than just quoting the number.
- Claude Code or Codex (or any AI chat, through the prompts at creatop.net/apollo-operator)
- The Apollo MCP connected (in Codex, the Apollo app connection)
- The Apollo CLI (
brew install apolloio/apollo-io-cli/apollo-io-cli) for bulk work and the kill switch - An email verification service. Pick your own; this library does not choose one for you.
Pick the way in that matches your tool. For Claude Code, the plugin is the one most people want.
In Claude Code, as a plugin (available in every session, from any directory, and updatable):
/plugin marketplace add creatop-gtm/apollo-operator
/plugin install apollo-operator@creatop
In Codex (the same 20 skills, adapted for Codex, plus a version of each for a chat without a terminal):
codex plugin marketplace add creatop-gtm/apollo-operator
codex plugin add apollo-operator@creatop
Start a new chat after installing so the skills load, and finish Apollo's sign-in in the app connection screen if it asks. The Codex plugin lives in plugins/apollo-operator/ and is rendered from the same source as the Claude skills.
In ChatGPT or any other chat: copy the start-here prompt from https://creatop.net/apollo-operator. It works out where you are and reads each step's method from this repo.
As your own skills (no plugin machinery, same result):
git clone https://github.com/creatop-gtm/apollo-operator
cp -R apollo-operator/skills/* ~/.claude/skills/
By opening the repo (to read the source, or to try it without installing anything): clone it and open the folder in Claude Code. The skills in .claude/skills/ load for that project only.
skills/ and .claude/skills/ hold the same twenty skills in the two layouts Claude Code reads, both generated from one source. Never hand-edit either.
Describe what you want ("help me define my ICP and build a list on Apollo") and operator-context routes you to the right phase.
Read .claude/skills/operator-context/references/outbound-principles.md. The short version: outbound is a system, relevance beats volume, deliverability is sacred, humans approve the copy, and these skills surface risks but never block. You decide.
Founders and go-to-market teams running outbound on Apollo who want a system, not a black box. Use it free to run your own motion. If you would rather have it built and run for you, that is what we do at Creatop.