MindEvolve.AIPhase 1 built · core technology

See how your agents are actually wired.

Point MindEvolve at an agent repository. It maps every agent, tool, skill, model and MCP server, and how control passes between them, with the source lines behind each connection. Your code is read, never run.

Become a design partnerHow it reads
9 frameworksstatic only · code never executedevidence on every edgeJSON · Mermaid export
MindEvolve Lattice mark
KEY FEATURE 01

Names the pattern, draws the architecture

MindEvolve detects the orchestration pattern (manager-worker, sequential or handoff) and draws it like an architecture document, with each agent's model, tools and skills. Click any arrow to see the source lines behind it.

manager-worker · sequential · handoffclick an edge, see the linesSVG · Mermaid · JSON
KEY FEATURE 02

Any branch, tag or commit

Pick a branch and MindEvolve analyses it without touching your working copy. Each branch becomes its own entry, so you can see how the architecture changed between them before it ships.

branch pickerworking tree untouchedone entry per branch
§1 · the problem

Nobody can draw the agent system they're running.

Agents get wired across frameworks, config files, prompts and MCP servers by different people at different times. Before you can improve an agent system, you need an honest picture of what it is. That picture is what MindEvolve produces first.

§2 · how it reads a repository

Three layers. Each one only adds.

fig. 1 — three-layer extraction into one architecture graph
L1 · STRUCTURE

Languages, dependencies, entrypoints, imports, prompts and agent config files. Every repo gets this skeleton.

L2 · DETECTORS

One plugin per framework reads agents, tools and handoffs from the code itself. Deterministic, and never overruled.

L3 · SEMANTIC

A model pass finds dynamic wiring the detectors can't. Anything it adds without a valid file and line is dropped.

OUTPUT

One architecture graph. Every node and edge carries its source layer, confidence and evidence.

every fact carries source, confidence, evidencemodel items without a valid file:line are droppeddeterministic facts are never deleted
§3 · what it maps

Every part, and every handoff

agentstoolsskillsmodelsMCP serverspromptsmemoryexternal APIsservicesentrypoints

Views for orchestration pattern, system layout and the raw graph. Click any edge to see the lines it came from. Export as JSON or Mermaid.

§4 · frameworks it reads

Whatever your team chose

Claude Agent SDKagents · skills · MCP
Anthropic SDK tool loopstools · models
LangGraphnodes · edges
CrewAIcrews · tasks
OpenAI Agents SDKhandoffs
AutoGengroup chat
Google ADKsub-agents
MCP serversexposed tools
Custom frameworkslow-confidence fallback
§5 · your code

Read, never run.

No execution

Customer code is never installed, imported or executed during mapping.

Secrets redacted

A secret scan runs on ingest. Values are redacted before anything is stored or sent to a model.

Model pass optional

Turn the semantic layer off and the analysis stays entirely deterministic and on your side.

§6 · roadmap

The map is the foundation.

PHASE 1 · MAPBuilt

Static analysis into an evidence-backed architecture graph, with a browser explorer and exports.

PHASE 2 · RUNNext

A sandbox built from the graph runs your system on our infrastructure and records traces and metrics.

PHASE 3 · EVOLVEPlanned

Monitoring and proposed improvements to prompts, tools, topology and cost, including new skills drawn from current research. Your team approves each one.

Design partners

Running agents in production? Send us a repository.

We'll map it with you and shape the next two phases around what you find.

Start a conversation