Code Mode Without the Code

by Bob Dickinson

Day 1 · Astor Ballroom (7th Floor) · MCP Best Practices · 25 min · 3 min read

Applied AI

TL;DR

Bob Dickinson presented MCP Graph, a no-code YAML-based framework for orchestrating MCP tool workflows. Instead of having agents write MCP server code, MCP Graph uses declarative YAML with JSON ADA for data transformation and JSON Logic for conditional routing. The project includes a visualization UI and a toolkit that lets agents build and test graphs through a devrail cycle.

Why Code Mode Creates New Problems

▶ Watch (3:04)

Three papers from late 2024 (Anthropic and others) advocated having agents build MCP servers that orchestrate tool calls in code. Two problems drove this: dynamic orchestration and context overload. Tool catalog bloat is largely solved, but tool calling bloat remains. Each tool call, its parameters, and its output stay in the context stack. A single document can add 100k tokens of irrelevant content. Subagents help but still burn tokens. Dickinson argues the orchestration problem is the more interesting part of the working set.

YAML Graphs Replace Generated Code

▶ Watch (9:23)

MCP Graph defines new tools as directed graphs in a single YAML file. The YAML uses the same shapes as the MCP spec for servers, imports, and tools. Five node types handle entry, MCP tool calls, data transforms, conditional routing, and exit. JSON ADA handles data transformation deterministically. JSON Logic provides conditional branching. The MCP Graph UX visualizes graphs, supports breakpoints, single-stepping, and context inspection. A 50-line YAML file can create a virtual MCP server that exposes only readonly tools with structured output.

Agent-First Tooling Through a Devrail Cycle

▶ Watch (15:40)

Dickinson’s first agents failed to build MCP Graph workflows because they lacked context he had as a human. His solution was the MCP Graph Toolkit, an MCP server that gives agents tools to inspect graphs, discover available servers, test JSON ADA transforms, single-step through execution, and deploy finished tools. He ran a devrail cycle: prompt an agent, watch it fail, ask it what tools would help, then iterate. Six cycles in 90 minutes produced agents that build graphs successfully every time.

Q&A

Was it an easy choice between JSON ADA and similar tools like JQ? Dickinson tested several models and found better support for JSON ADA based on training data availability. ▶ Watch (21:08)

Does MCP Graph support parallel execution and full DAGs? The current version handles sequential workflows with switches but lacks map/reduce primitives, though Dickinson said adding them is reasonable. ▶ Watch (21:55)

Does the toolkit use MCP annotations for tool capabilities like open world access? Not yet, but Dickinson noted it aligns with his observability goals for understanding what individual tools do. ▶ Watch (23:44)

Notable Quotes

executable code actions elicit better LLM agents Bob Dickinson · ▶ Watch (1:24)

this generated source code with access to your Salesforce and your Google Docs MTP servers Bob Dickinson · ▶ Watch (6:43)

I got amazing, like shockingly good feedback, like way better than I’ve ever gotten from a human doing this kind of thing Bob Dickinson · ▶ Watch (18:51)

Key Takeaways

MCP Graph replaces generated MCP server code with declarative YAML graphs.
The MCP Graph Toolkit lets agents build workflows through a devrail cycle.
JSON ADA and JSON Logic provide safe, deterministic data transformation.

About the Speaker

Bob Dickinson is a serial founder, CTO at scale, and always a hands-on builder. Creator of MCP Tool Vault and the open source projects tsAgent and mcpGraph. Maintainer of MCP Registry and MCP Inspector. Background in security, including as CTO of OneLogin and Censys.

Filed under

Applied AI