What is Elemm?

Introduction to the Landmark Manifest Protocol for autonomous AI agents.

The Problem: The Swiss Army Knife Paradox

Standard tool-calling protocols for AI agents (like classic MCP) provide the agent with all available tools at once. As the toolset grows (50+ tools), critical issues emerge:

Context Bloating — The system prompt becomes massive, consuming tokens and increasing costs dramatically.
Latency — Large prompts increase time-to-first-token significantly.
Hallucinations — The model loses focus, confuses similar tools, or ignores parameter constraints.

The Solution: Semantic Landmarks

Elemm solves this through a tiered discovery system called Landmarks:

Landmarks function as an interactive table of contents.
The agent initially only sees high-level namespaces (e.g., Security, IT, HR).
Detailed tool signatures are only loaded when the agent proactively "steps into" a specific landmark.
This keeps the agent's context clean and focuses its reasoning power on the relevant subset.

How Does It Work?

The protocol follows a five-stage handshake:

Connect — The agent connects via connect_to_site(url) to a remote API.
Manifest — The agent calls get_manifest() and receives protocol rules + the landmark topology. No tool signatures!
Landmarks — get_landmarks() shows available functional areas.
Inspection — The agent selects relevant landmarks and calls inspect_landmark(id) for full technical signatures.
Execution — Only now does the agent execute actions via call_action() or execute_sequence().

💡 TIP

Agents can also use search_landmarks(query) to directly search for specific tools via regex — without traversing the full hierarchy.

The Gateway: A Universal Broker

The Elemm Gateway is a universal MCP server that acts as a protocol-aware broker between AI agents and remote APIs.

Multi-Protocol — Seamlessly supports OpenAPI/Swagger, GraphQL, and native Elemm APIs.
Tool Isolation — Only 9 generic core tools are exposed to the agent. All domain-specific actions are routed through call_action and execute_sequence.
Security — A built-in Security Policy Engine (Guardian) enforces restrictions before every API access.

Benchmark Results

In standardized tests using 120B parameter models, the Landmark Manifest Protocol shows massive improvements over classic MCP:

Metric	Classic MCP	Elemm Protocol	Improvement
Avg. Steps	12.80	2.00	-84.4%
Input Tokens (Avg)	92,024	7,081	-92.3%
Output Tokens (Avg)	2,545	692	-72.8%
Avg. Duration	65.35s	23.72s	-63.7%
Estimated Cost	$2.15	$0.22	-89.5%

❗ IMPORTANT

Even with ultra-lightweight models (0.8B parameters), Elemm raises the success rate from 10% to 70% — a 600% improvement.

Why Token Efficiency Matters

Context Optimization

By isolating tool signatures within landmarks and only loading them on demand, the initial prompt size is reduced by over 90%.

Turn Consolidation

execute_sequence allows planning and executing complex logic chains in a single turn. What typically requires 12+ roundtrips is resolved in just 2 turns with Elemm.

Server-Side Piping

Native variable piping ($step0.id) eliminates the need for the LLM to manually extract and re-inject data.

Caching Advantage

Elemm's highly stable prompt prefix (System Prompt + Manifest) ensures a near 100% cache hit rate for tool definitions on providers with prompt-/context-caching.