Agentic Atlas MCP¶

Public, read-only consultation surface for field-tested agent-system design patterns - a hosted knowledge corpus that both people and agents can read, published by agentic-atlas.dev. Five keyless tools map ordinary design vocabulary to canonical patterns, read the corpus at card, section, or full-node depth, and traverse relationships between nodes. Live-probed Release v1.0.6, served stateless at https://agentic-atlas.dev/mcp/ (protocol 2025-06-18) - no account, no credentials.

Server type: Remote (Streamable HTTP, hosted, stateless)
Auth: None
Release: v1.0.6 (serverInfo; registry: dev.agentic-atlas/atlas)
Website: https://agentic-atlas.dev/
Endpoint: https://agentic-atlas.dev/mcp/
Tools: 5 (live-probed 2026-09-08; protocol 2025-06-18)
Pricing: free
Category: AI & Agents
Built by: Avery Jones (github.com/aj604/agentic-atlas-plugin)

Why This Matters for Operators¶

Agent systems fail in predictable ways - unbounded context, unclear ownership of decisions, drift between what was designed and what was shipped. Agentic Atlas is the reference layer for those failures: patterns that have been field-tested, with every card claim carrying its source address. The corpus is served as one active Release (v1.0.6): the endpoint serves exactly that release, and a promotion replaces it wholesale - no silent drift between what a caller read last week and what it reads now.

For a team building or buying agent infrastructure, this is a cheap, keyless way to have an agent consult established design practice before it invents one.

Tools & Capabilities¶

5 read-only tools confirmed by live probe of Release v1.0.6 (2026-09-08):

Tool	Purpose
atlas_orient	Entry point: discover canonical identity from an ordinary-language query (e.g. "context window budget") when no canonical id is known; ranks whole-word matches by how few subjects carry them
atlas_cards	Read 1-4 canonical ids as one ordered Card per id: identity, Hook, and sealed claims with source addresses
atlas_read	Read a complete Node - bare id: identity, hierarchy, Hook, and every Section
atlas_links	Page one subject's relationship occurrences, outbound then inbound, with direction and reference context
atlas_navigate	Read "tree" (canonical ordered traversal of every published node) or "tour" (publisher-curated path)

The former atlas_provenance, atlas_define, and atlas_decisions tools were retired in Release v1.0.6; their capabilities are recovered through this five-tool grammar (provenance and publication decisions surface as source addresses on sealed card claims; glossary terms resolve through atlas_orient vocabulary matching). The endpoint also publishes a consultation contract and three companion workflow skills (designing-agent-systems, reviewing-agent-designs, applying-a-pattern) as resources for hosts implementing the draft MCP Skills Extension.

Calls are idempotent and read-only (annotated readOnlyHint), and results carry scope metadata (total/returned/truncated) with cursor pagination. An empty result is a successful empty payload, not an error.

Installation¶

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

No local installation, no account, no credentials. Works with Claude Code, Cursor, VS Code, Codex and any MCP-compatible client.

Configuration¶

{
 "mcpServers": {
 "agentic-atlas": {
 "type": "http",
 "url": "https://agentic-atlas.dev/mcp/"
 }
 }
}

The server is stateless: initialize returned no session ID and tools/list succeeds without one (MCP '26-style stateless transport). No API key of any kind.

Business Relevance¶

Operators building agent systems get a consultation surface for design decisions - when to add memory, how to structure approval gates, which patterns have held up in the field.
Architects reviewing agent stacks get a source address on every card claim, so a design choice can be traced to its published evidence.
Teams onboarding agents get a stable vocabulary and canonical naming layer that a new agent can read on first contact.

Integration with CorpusIQ¶

CorpusIQ is the business-data layer (QuickBooks, Stripe, HubSpot, GA4 and 40+ more connectors); Agentic Atlas is the design-practice layer. A Hermes agent composing a workflow can consult Agentic Atlas for the pattern (how should approval gates and read-only contracts be structured?) and then execute against CorpusIQ for the data (what do the numbers say?). Design wisdom from Agentic Atlas, business truth from CorpusIQ.

Limitations¶

New listing (added Aug 24, 2026; refreshed Sep 8, 2026 for Release v1.0.6) - no long-term track record yet.
Read-only consultation surface: it provides design guidance, not transactional APIs or execution.
Single-source corpus: patterns reflect the publisher's curation and field testing, not an open contribution pool.
No authentication means no personalization - every caller reads the same active Release.