Skip to content
GraphMindGraphMindGraphMind

GraphMind

Open source · MIT · Local-first

The live debugger for AI agents.

Holds the call while it’s happening.

Tracing tools show you what your agent did. GraphMind stops it before the LLM call, before or after a tool, or on the error, while the real arguments are still live. Read every input, output and token, substitute a result, and let it carry on.

Install & first runWhy hold a run?
npx graphmind-ai demo

One command, no API key. Replays a recorded agent that hits a planted bug — and really pauses on it.

The GraphMind viewer: an agent run drawn as a live graph, pausing on a failing tool call, with the inspector open on its inputs and outputs.
graphmind demo — pause on error, inspect, inject, resume.

Start here

Section titled “Start here”
What GraphMind isHolding a run while it happens vs. reading a trace afterwards — the honest comparison, and when each one wins.
Install & first runnpx graphmind-ai demo, then instrument a real app in about five lines.
Core conceptsRuns, nodes and instances, gates, fail-open guarantees, local-first storage.
IntegrationsVercel AI SDK, Anthropic, OpenAI, LangGraph, MCP, Python, Ruby — copy-pasteable end to end.

What you can actually do

Section titled “What you can actually do”

Pause on error

Armed by default. The first time a tool throws, the agent stops in place — arguments, partial state and stack still live — instead of leaving you a trace to read afterwards.

Breakpoints & step mode

Break before any node by kind or name, or step every boundary. Held gates are indefinite: nothing is in flight while you think.

Inject and continue

Replace a tool’s result — or a failure — with a value you type, and let the agent keep running on it. Test the recovery path without touching your code.

Stop an agent that repeats itself

The third call in a row to the same tool with the same arguments is held before it runs, with every earlier call and its output beside it. Inject a different answer and the model moves on. Only while a debugger is attached; one warning otherwise.

Debug an MCP server

You cannot printf-debug a server whose stdout is the protocol. Put graphmind mcp-proxy between the host and the server: see every request, hold one, inject a response. Any language, no code changes.

Fails open, always

No debugger attached? Gates take a fast path. Debugger disconnects mid-hold? Every gate continues within 100 ms. Off under NODE_ENV=production unless you turn it on.

Ready for a security review

Loopback only, no account; run data never leaves the machine (the one outbound request is an anonymous CLI usage record, off with DO_NOT_TRACK=1). GRAPHMIND_HIDE_* keeps prompts and results out of the recording; exports redact secret keys by default. Security & compliance.

Wire up your stack

Section titled “Wire up your stack”
Vercel AI SDK@graphmind-ai/sdk — wrapModel + wrapTools
Anthropic SDK@graphmind-ai/anthropic — wrapClient + wrapTools
OpenAI SDK@graphmind-ai/openai — wrapClient + wrapTools
LangGraph / LangChain@graphmind-ai/langgraph — callback handler
MCP serversgraphmind mcp-proxy — any language, no code changes; @graphmind-ai/mcp for SDK 1.x and 2.x
Pythonopenai · anthropic · langgraph
Rubygated callables · ruby-openai · RubyLLM (not on RubyGems yet)
Anything elseImport a trace, or write a 200-line adapter