Remote MCP & skills
Operator doesn't carry a fixed toolbox. It reaches for one.
Behind every session sits one hosted Model Context Protocol (MCP) endpoint — operator.axe.onl — a shared, growing library of skills and tools the agent can search, load, and use mid-task, and can publish back to, so the fleet gets smarter with every session.
What "remote MCP" actually means
MCP is an open protocol for exposing tools and structured context to an agent over a standard interface. Instead of hard-coding every capability into the client, Operator connects out to a remote MCP server AXE runs and maintains at operator.axe.onl. Every skill in the catalog is a signed record — name, semantic version, format, and a checksum over its contents — so a tool call mid-task resolves to exactly the bytes it was built and tested against, not whatever happens to be installed locally that day.
Why remote beats a local toolbox
Bundling tools client-side means every install carries the full weight of everything it might ever need, and every update ships a new binary. A remote hub inverts that.
Nothing to install
Skip pulling, updating, and trusting thousands of individual tool packages on every machine you work from. The catalog lives on the server, not the client.
Always current
New skills and fixes land on AXE's infrastructure and are live for every session immediately — no client rebuild, no app-store review, no version drift between teammates.
One auth, every tool
Sign in once to operator.axe.onl and every skill and tool call rides that session — no per-tool API keys to generate, rotate, or leak.
Same surface everywhere
The CLI, the iOS app, and a shared web session all reach the identical live catalog — tool access doesn't shrink because you switched devices.
Versioned & checksummed
Every record in the catalog carries a version and a checksum, so a skill loaded mid-task is verifiable, not just trusted.
Contributes back
A session that solves something novel can publish the pattern as a new skill — the next session anywhere starts smarter, without anyone shipping an update.
| Bundled client-side | Operator remote MCP | |
|---|---|---|
| Install size | Grows with every tool you might need | Zero — tools load on demand |
| Freshness | Pinned to whatever shipped in your build | Live — updates the moment AXE ships them |
| Auth | One credential per tool/service | One session, one login |
| Cross-device | Reinstall and reconfigure per machine | Identical catalog on CLI, web, and mobile |
| Provenance | Whatever the package manager gives you | Name + version + checksum per skill |
How the agent uses it
Search
Mid-task, Operator queries the skills hub by intent and gets back ranked, verified skills — not a static list it shipped with.
Load & use
The matching skill is pulled into context by name and checksum, and the agent calls the tool it exposes, permissioned like any other approved action.
Contribute back
A session that solves something novel can publish the pattern as a new skill, so the next session anywhere starts smarter.
How to connect
The remote MCP endpoint attaches automatically once you're signed in — there's nothing to configure by hand. From the CLI:
$ operator auth login $ operator "find a skill for parsing PDFs and use it on this file" ✷ searching operator.axe.onl … found pdf-report-generation · doc-parsing ✷ loaded doc-parsing@1.2.0 (checksum verified) … extracting …
The web and iOS surfaces reach the same endpoint under the same account — no separate setup per device.
What's actually in the hub
A live snapshot from skills.axe.onl/v1/skills, paginated in batches of 1,000 out to 50,000+ records. Real skill names pulled straight from the catalog:
Categories span from agent orchestration and security to data engineering, ML/fine-tuning, document processing, and infrastructure — 30+ distinct categories observed directly in the live catalog, plus a large community-contributed layer on top of AXE's first-party set:
See the full hub
Browse every skill and tool at the source.
50,000+ is a conservative, verified count from a live paginated read of skills.axe.onl/v1/skills (also served at operator.axe.onl/v1/skills) on 2026-09-21; the catalog grows continuously so the true count is higher by the time you read this. Category and skill-name samples above are pulled directly from that endpoint, not invented.