GraphMind is out — open source, MIT, local-first. Star it on GitHub.

GraphMind

Open source · MIT · Local-first

The debugger for AI agents.

Attaches while it’s happening.

Your traces shouldn’t be a post-mortem. GraphMind attaches to the run itself: watch your agent execute as a live graph, pause on error, set breakpoints, step through every input and output, inject a fix, and resume.

npx graphmind-ai demo

One command, no API key — watch a real debug session pause on a planted bug. Or try the debugger in your browser, nothing to install.

Occasional release notes as GraphMind grows. No spam, just ships.

support-agent AGENT step ×3 LLM Converts to $60,000,000. Checking against budget. 184 → 52 tok 581ms searchFlights DONE 312ms getWeather DONE 88ms checkBudget TOOL BudgetExceededError paused on error

What it does

01

Live attach

Attach to a running workflow and watch it execute as a graph — every node, edge, and tool call, while it happens. Not a trace viewer for afterwards.

02

Pause & step

Set breakpoints on any node. Pause automatically on error. Step through the run the way you step through code.

03

Inspect & inject

Open any node and read its exact inputs and outputs. Edit state, inject a fix, and resume — no restart, no replay.

04

Local-first, MIT

npx graphmind-ai and you’re debugging. Loopback-only server, keys kept out of the recording, run data that stays put. MIT licensed.

How it works

The graph is your code.

GraphMind never authors your workflow — the graph is a projection of the code you already wrote. Wrap your model and tools, run your agent, and the debugger attaches. There are adapters for the Vercel AI SDK, the Anthropic and OpenAI SDKs, LangGraph/LangChain and Python, plus generic OpenTelemetry and OpenInference trace import — and a proxy that debugs an MCP server in any language with no code changes at all.

  1. npx graphmind-ai init detects your framework and prints the snippet to add.
  2. npx graphmind-ai starts the local debugger.
  3. Run your agent. The graph lights up: breakpoints, stepping, inject-and-resume, live state.

Need to hand a bug to someone else? npx graphmind-ai record <run> --html writes the whole run to a single self-contained HTML file — it opens in any browser, with no server and no account.

Vercel AI SDK agent.ts 
import { graphmind } from '@graphmind-ai/sdk';

const gm = graphmind({ app: 'support-agent' });

const model = gm.wrapModel(anthropic('claude-sonnet-4-5'));
const tools = gm.wrapTools({ searchFlights, checkBudget });

Local-first, precisely: the server binds 127.0.0.1 only and refuses connections from foreign browser origins. Your API keys never enter the recording — a dedicated credential-leak audit asserts it on every build. Your prompts, outputs and traces stay on your machine. The CLI does send one anonymous ping per command (its name, a random install id, the version — never arguments, prompts or run data), which GRAPHMIND_TELEMETRY=0 turns off; here is exactly what it sends.

Works with your stack

Instrument what you already wrote.

Not sure which you need? npx graphmind-ai init reads your project and tells you.

For teams

The debugger is free forever. Teams get more.

A self-hosted Team edition is in the works: shared run history, workspaces, and SSO, running entirely on your own infrastructure with flat annual pricing. No per-seat metering, no traces leaving your network.

Want it early, or want to shape it? Email hello@graphmind.ai — design partners get it first, at a permanent discount.