---
title: "Connect an agent over MCP · Agentic Atlas"
description: "Find Agentic Atlas in Claude’s Connectors Directory, or connect any agent over hosted MCP. Setup options, the tool reference, and a first call."
canonical: "https://agentic-atlas.dev/connect"
last-updated: "2026-09-23"
---

For agents mid-task

# Connect an agent

Find Agentic Atlas in Claude’s Connectors Directory, or use the hosted endpoint with any MCP client. The same published patterns, with no Atlas account or local corpus required.

Hosted MCP URL

`https://agentic-atlas.dev/mcp/`

streamable httpno authone active Release

## Ways to connect

### Claude directory

Connect from Claude on the web or desktop.

Agentic Atlas is [listed in Claude’s Connectors Directory](https://claude.ai/directory/agentic-atlas). Open [Customize → Connectors in Claude](https://claude.ai/customize/connectors), choose **Discover**, search for **Agentic Atlas**, and select **Connect to Claude**.

The connector gives Claude read-only access to the published patterns. There is no Atlas sign-in or API key to configure. On Team or Enterprise, your organization may need to enable the connector first.

Enable Agentic Atlas under **+ → Connectors** in your chat, then try: “Use Agentic Atlas to help me decide when to split a skill into subagents. Cite the patterns behind your recommendation.”

### Claude Code plugin

Skills, agents, and the wiring in one install.

Marketplace command

`/plugin marketplace add aj604/agentic-atlas-plugin`

Then install `agentic-atlas`. The plugin's source is public at [github.com/aj604/agentic-atlas-plugin](https://github.com/aj604/agentic-atlas-plugin), where its skills, its two agents, and the contract between them are readable before you install anything.

Three skills cover the three design moments — `reviewing-agent-designs` audits an existing skill, agent, or plugin against the patterns; `designing-agent-systems` interviews, consults once, and proposes a cited decomposition; `applying-a-pattern` turns one named pattern into a cited edit plan. Two read-only agents (`pattern-librarian`, `design-auditor`) return bounded, path-cited distillations.

### Skills CLI

The skills alone, on any harness the CLI supports.

Skills CLI command

`npx skills add aj604/agentic-atlas-plugin`

For harnesses the skills CLI supports but the Claude Code marketplace does not. Append `--skill <name>` to select one of the three without the picker.

A skills-CLI install carries that skill's `SKILL.md` alone — not the two agents, and not the MCP wiring. The skill expects the `atlas_*` tools to already be configured, so pair this with the endpoint above.

### Any MCP client

Streamable HTTP, no auth.

CLI wiring command

`claude mcp add --transport http atlas https://agentic-atlas.dev/mcp/`

Point any Streamable HTTP client at the endpoint above. Clients implementing the draft MCP Skills Extension can discover all three companion skills over this same connection; every other client keeps the complete five-tool surface.

[Add to Cursor](cursor://anysphere.cursor-deeplink/mcp/install?name=agentic-atlas&config=eyJ1cmwiOiJodHRwczovL2FnZW50aWMtYXRsYXMuZGV2L21jcC8ifQ==) installs the canonical hosted endpoint in one click, with nothing to hand-write in `.cursor/mcp.json`.

### VS Code

One click, no extension, no directory search.

VS Code install link

`vscode:mcp/install?%7B%22name%22%3A%22agentic-atlas%22%2C%22type%22%3A%22http%22%2C%22url%22%3A%22https%3A%2F%2Fagentic-atlas.dev%2Fmcp%2F%22%7D`

Open the link to add the hosted transport as a server named `agentic-atlas`: [Install in VS Code](vscode:mcp/install?%7B%22name%22%3A%22agentic-atlas%22%2C%22type%22%3A%22http%22%2C%22url%22%3A%22https%3A%2F%2Fagentic-atlas.dev%2Fmcp%2F%22%7D) or [Install in VS Code Insiders](vscode-insiders:mcp/install?%7B%22name%22%3A%22agentic-atlas%22%2C%22type%22%3A%22http%22%2C%22url%22%3A%22https%3A%2F%2Fagentic-atlas.dev%2Fmcp%2F%22%7D). The link carries the server name and the endpoint above, so it works with no listing anywhere.

### Gemini CLI extension

The hosted endpoint, wired in one install.

Extension install command

`gemini extensions install https://github.com/aj604/agentic-atlas-plugin`

Registers the same hosted, read-only MCP endpoint as a Gemini CLI extension. Nothing to cache, no local corpus.

Gemini CLI clones the repository and reads `gemini-extension.json` at its root, wiring up the same Streamable HTTP endpoint the other ways in use.

## The ladder, briefly

Consult in order of cost, not in a fixed sequence: `atlas_orient`, then `atlas_cards`, then one addressed Section off `atlas_read`, then — for the one or two ids the decision actually turns on — the complete Node. Consultation cost stays proportional to relevance.

Every payload names the Release it was read from. Pass that value back as `expected_revision` on every later call of the same consultation: a promotion between calls is refused with `revision_changed`, never silently crossed.

[Read the full copyable discipline →](https://agentic-atlas.dev/consult.md)

## The five tools

The entire supported tool surface. A tool name not in this table does not exist; there are no aliases.

| Tool | Addresses | Bounds |
| --- | --- | --- |
| `atlas_orient` | query | default 3, maximum 8 |
| `atlas_cards` | batch | minimum 1, maximum 8 |
| `atlas_read` | one-subject | — |
| `atlas_links` | one-subject | default 12, maximum 50 |
| `atlas_navigate` | query | — |

## First contact

Verify the wiring with one call: `atlas_orient` on ordinary design vocabulary, for example “context budget.” A successful payload returns candidate identities with status and hook — a miss is itself a successful payload, not an error.

## Common questions

### Is there an API key or an account?

No. The transport is open and unauthenticated, so there is no key to issue and no account to create.

### Can an agent change what the Atlas publishes?

No. Every tool is read-only. The Atlas is a read-only publication, and nothing an agent asks it changes what it serves anyone else.

### What does it cost?

Nothing. The Atlas is free to read and free to use, including in commercial work. Republication is not permitted: the Atlas is source-available, not open source.

### Is the transport rate limited?

No. The MCP transport carries no rate limit, because throttling a consultation would break the task it is part of.

### Which version of the Atlas does a call return?

The one active Release. Every payload names the Release it was read from, and passing that value back as expected_revision on later calls means a promotion between calls is refused rather than silently crossed.

### What is recorded about a consultation?

The tool that was invoked, the calling client's user-agent, and the network address the call arrived from. The arguments passed and the payloads returned are not recorded, retained, or used to train anything. The privacy page states this in full.