Runtime agents

Audience: AI agents with (or about to use) the feel-your-protocol MCP server — not human onboarding, not server development.

Source of truth

When MCP is connected, trust the live server first:

1. listTools — tool names, descriptions, and input schemas
2. describe_capabilities — baselineForkId, forks (osaka baseline, amsterdam preview), runnable EIP modules, opcodes, encoding, ceilings

Markdown on this site is secondary. It can lag behind a gateway release; the probe response and tool schemas cannot.

If MCP is connected and working, you usually do not need this page — unless the user pasted a doc URL or you need behavior rules below.

Not this page

If you are…	Go to…
A human connecting or exploring	Connect, EIP catalogue, EIP-8024
Extending the engine or gateway	Contributing → mcp-execution-engine AGENTS.md

Calling tools

1. Probe first — learn what is runnable and how opcodes encode
2. Run with caller-supplied bytecode on the fork you need (default amsterdam). Optionally run again on osaka when comparing against mainnet — this server does not ship demo programs
3. Do not substitute the local mcp-execution-engine lab, npm run lab, or repository source unless MCP is unavailable

Replying to humans

When the user is a human (exploring protocol changes, not integrating the server):

• Answer in plain language about behavior and results (gas, success, stack, traces)
• Call MCP tools silently
• Do not expose tool names, JSON field names (such as baseHardfork), raw hex, or request payloads unless the user explicitly asks for implementation detail

Reporting results

Always cite provenance.engineVersion and provenance.forkConfig when reporting simulation outcomes.

When you are reading docs (no MCP)

Use this order:

1. Capabilities — query shapes
2. Per-tool reference: Discover, Run
3. Guarantees — determinism, provenance, ceilings

Machine-readable index: /llms.txt

Full use-layer text: /llms-full.txt

JSON schemas: describe_capabilities, run_evm_bytecode

Changelog

Runtime Agents Changelog

1. v0.92026-08-27compare_evm_variants removed — simulate twice to diff.
2. v0.82026-08-27Narrow scope — MCP-first runtime playbook; renamed from for-ai-agents; builders pointed to engine AGENTS.md.
3. v0.72026-08-27Replying to humans — plain language; hide tool names and JSON fields unless asked.
4. v0.62026-08-27Agents construct bytecode; catalog exposes opcodes/encoding, not demo programs.
5. v0.42026-07-22Live MCP tools, schema URLs, explicit guidance to prefer MCP over local lab.
6. v0.32026-07-20Initial AI/LLM reader guide under use/.