VisualAgents

A live map of which agent in your multi-agent AI system is doing what — built from its chain of thought.

Multi-agent apps usually show you one long stream of thoughts and tool calls. You can't tell which specialist is working, where the request is in the workflow, or why it just looped back. VisualAgents reads your agent project, draws its agents and handoffs as a map, and lights that map up live while the agent works — from the terminal, on your own machine, with no changes to your code.

Contents · Highlights · Quick start · Watching an agent live · The dashboard · How it works · Supported frameworks · Choosing the LLM · Instrument anything · Examples · CLI reference · Privacy · Troubleshooting

Highlights

• Maps any agent repo. Static analysis understands LangGraph, CrewAI, AutoGen, the OpenAI Agents SDK, Swarm, Google ADK, smolagents, LlamaIndex and raw OpenAI-SDK apps; an optional LLM pass (Claude, GPT, Gemini or a local model) adds roles, tools, handoffs and the step-by-step workflow.
• Live map. The working agent glows with a thought bubble, handoffs animate along the arrows (with ×N counts for loops), workflow steps tick off, and the chain of thought streams beside it — colour-coded per agent.
• Rewind and replay. A time ruler shows every agent's turn. Drag back to any second and the whole dashboard rewinds; press play to replay at up to 16×. Every run is recorded and can be replayed later.
• No setup per project. Run visualagents and answer three questions. It finds how the project starts, uses the project's own virtualenv, and remembers your choices. Or hook a folder once and keep running your agent however you like.
• Precise where it can be, smart where it can't. Auto-instrumentation gives exact agent boundaries for LangChain/LangGraph, the OpenAI Agents SDK and direct OpenAI-SDK calls; everything else is attributed from the console output, with optional LLM narration.
• Interactive. Drag blocks around (arrows follow), hover a line of thought to make its agent bounce, click an agent for its own history, switch light/dark, share a link to an exact moment.
• Private by default. Everything runs on 127.0.0.1. Your code only goes to an LLM if you pick one, with secrets redacted first.

Quick start

1. Install

Requires Python 3.10+.

uv tool install "visualagents[all] @ git+https://github.com/kaih06-hub/VisualAgents.git"

or with pipx: pipx install "visualagents[all] @ git+https://github.com/kaih06-hub/VisualAgents.git". Claude support is built in; [all] adds Gemini and OpenAI / OpenAI-compatible servers (Ollama, LM Studio, vLLM). From a local clone: uv tool install --editable ./VisualAgents --with google-genai --with openai.

2. See it work (no API key)

visualagents demo

A simulated four-agent research crew — planner → researcher → writer ⇄ critic — runs in your terminal while the dashboard maps it.

3. Point it at your own agent

visualagents

Three questions, then your agents' map pops up:

  VisualAgents  · a live map of your agents
  Three questions, then your agents' map pops up. Ctrl+C to quit.

1/3 Which project should VisualAgents map?
  ✔ ~/VisualAgents/examples/langgraph_supervisor

2/3 Which LLM should read the code and draw the map?
  ✔ Google Gemini
  Model
  ✔ gemini-2.5-flash

3/3 Google Gemini API key
  Paste it below - it stays hidden · get one at https://aistudio.google.com/apikey · empty = skip
  API key:
  ✔ Key works · 34 models available
  ✔ Saved to ~/.visualagents/keys.json

  ✔ Found 4 agent(s): Supervisor, Researcher, Coder, Reviewer  (langgraph)
  ✔ Map is open at http://127.0.0.1:7420

▶ Watch it live?
  ❯ Start it now  python app.py
    Start it with a different command…
    Hook it - I'll start it myself  IDE, scripts, notebooks: it shows up whenever it runs
    Just show the map

1. Project — a recent one, the current folder, any path (Tab completes), or a GitHub URL, which it clones.
2. LLM — Gemini, Claude, OpenAI, a local / OpenAI-compatible server, or offline; then a model (best / cheaper / cheapest, or any id).
3. API key — pasted hidden, checked with a free model-list call, and (if you agree) saved to ~/.visualagents/keys.json, readable only by you. Next time you just press Enter through the defaults.

Menus work with the arrow keys (or j/k and numbers).

Watching an agent live

The map moves only while the agent is actually working. VisualAgents watches; it doesn't do the work. So:

1. Start the agent — choose Start it now in the terminal, or press ▶ Start on the map. VisualAgents runs the project's start command in your terminal.
2. Give it a task, the way that agent takes input:
   • agents that ask questions in the terminal (e.g. TradingAgents asks for a ticker) → answer them there;
   • web-app agents (e.g. LAMBDA) → open the agent's own page and type your request;
   • agents that take the task as an argument → put {task} in the start command (visualagents run . -- python main.py "{task}") and a What should it do? box appears on the dashboard.
3. Watch the map light up. When the run ends, ▶ Run again starts the next one.

There are three ways to connect VisualAgents to an agent — pick whichever fits how you work:

visualagents run ~/projects/my-agent --pick                 # choose the start command again
visualagents run ~/projects/my-agent -- python main.py "q"  # or give an exact command
visualagents hook ~/projects/my-agent                       # then: ./start.sh, your IDE, …
visualagents status | open <folder> | unhook <folder> | up | down

Start scripts that push the app into the background or send its output to a log file still work: events reach the dashboard over HTTP, and *.log files the app writes are followed. You can also feed it output you already have: python main.py 2>&1 | visualagents pipe ., or visualagents watch . --log run.log.

The dashboard

Live map

• Graph — agents laid out in workflow order (left→right, or top→bottom on narrow screens). The working agent glows with a bubble showing what it's thinking; finished agents get a ✓; arrows show who hands work to whom, with traversal counts. Two-way handoffs are drawn as parallel arrows; handoffs that weren't in the blueprint appear dotted, and agents it missed appear dashed.
• Steps — the blueprint's phases across the top, checked off as the run moves through them.
• Working now — the active agent, its current activity, time in this turn, tool calls and handoffs.
• Live chain of thought — every thought, tool call, result and answer, colour-coded per agent; filter by agent or hide plain logs. Hover any line and its agent lights up and bounces on the map.
• Drag any block to rearrange the map — arrows follow, positions are remembered per project, and the layout button restores the automatic layout. Click a block for that agent's details and its own chain of thought.

Blueprint

The first time you open a project the blueprint pops up (press B any time): a summary, where requests enter and which agent answers, the step-by-step flow, every agent with its role, tools and source location, and every handoff. It scrolls in one column, top to bottom.

Time ruler — rewind and replay

Each agent's turn is a segment on the ruler, with the workflow steps above it. Drag the playhead back and the whole dashboard rewinds — who was working, which step, what it was thinking, and the chain of thought up to that second (later lines dim). ▶ / Space replays at 1–16×; Live / L jumps back to now.

Dark mode and small screens

Light, dark or follow the system (button in the top right). On narrow screens the map flows top-to-bottom.

Shortcuts and links

How it works

flowchart LR
  repo["Your agent project"] -->|static analysis + optional LLM| bp["Blueprint<br/>.visualagents/blueprint.json"]
  repo -->|run · hook · pipe · watch| proc["Agent process"]
  proc -->|framework hooks, OpenAI SDK hook,<br/>console output, log files| engine["Attribution engine"]
  bp --> engine
  engine -->|Server-Sent Events| ui["Dashboard on 127.0.0.1"]
  engine -->|every input| rec[".visualagents/runs/*.jsonl<br/>(replayable)"]

1. Blueprint. visualagents scan (run for you on first use) reads the most agent-relevant files. Static analysis finds agents, tools and handoffs for the frameworks below; the optional LLM pass returns each agent's role, description, tools, source location, log aliases and keywords, the handoff graph, entry/exit agents and the general workflow, as schema-checked JSON. You can commit or hand-edit the result.

2. Attribution. While the agent runs, three sources say which agent is doing what, from most to least precise:

   Source	When	What you get
   Auto-instrumentation	Python apps using LangChain / LangGraph, the OpenAI Agents SDK, or the OpenAI SDK directly (any OpenAI-compatible endpoint; sync, async, streaming)	Exact agent boundaries, LLM replies as thoughts, every tool call and result — no code changes. Direct SDK calls are credited to the class or function that made them (e.g. Programmer, Inspector). Turn off with --no-instrument.
   Console patterns	Always	CrewAI (# Agent: …, rich panels), AutoGen (coder (to critic):), LangGraph stream output, ReAct (Thought: / Action: / Observation: / Final Answer:), handoff phrases (Handing off to Writer, transfer_to_billing), [Agent] … prefixes, and tools owned by one agent.
   LLM narration (run --smart)	Optional	Every few seconds an LLM reads the latest output and names the working agent with a one-line summary.

3. Dashboard. A small local web server streams the state to the browser; the map, steps, feed and ruler are rebuilt from the event log, which is what makes rewinding possible.

Supported frameworks

Choosing the LLM that draws the map

visualagents models                                   # which providers have keys, and their defaults
visualagents models --provider google                 # every model your account can use
visualagents scan . --provider google --model gemini-2.5-flash
visualagents scan . --static                          # offline only

Keys and choices are read from, in order: your shell environment, the project's .env, then what the guided start saved. You can also set VISUALAGENTS_PROVIDER, VISUALAGENTS_MODEL and VISUALAGENTS_NARRATE_MODEL. Gemini requests retry automatically when the model is busy.

Instrument anything yourself

From any language, print a line:

@@visualagents {"type": "agent_start", "agent": "researcher", "text": "find sources"}

From Python:

from visualagents import agent, emit, handoff, track

with agent("researcher", task="find sources"):
    emit("thought", "researcher", "Start with peer-reviewed numbers")
    emit("tool_start", "researcher", "solid-state battery Wh/kg", tool="web_search")
handoff("researcher", "writer", "findings ready")

@track("critic")                      # the whole function is the critic's turn
def review(draft): ...

Event types: agent_start, agent_end, thought, tool_start, tool_end, handoff, output, log, error. Under visualagents run the dependency-free runtime is importable even if VisualAgents isn't installed in your app's environment (from visualagents_runtime import agent, emit), and marker lines are hidden from your terminal. Without run: LangChain → callbacks=[visualagents.langchain_handler()]; OpenAI Agents SDK → agents.add_trace_processor(visualagents.openai_agents_processor()).

Examples

LAMBDA (a data-science agent). LAMBDA is a FastAPI backend plus a React UI that calls the OpenAI SDK directly, and its start.sh runs both in the background:

git clone https://github.com/AMA-CMFAI/LAMBDA.git ~/LAMBDA
cp ~/LAMBDA/backend/.env.example ~/LAMBDA/backend/.env    # add its model key
visualagents run ~/LAMBDA                                  # pick ./start.sh the first time

Ask LAMBDA something at http://localhost:3000 and watch it at http://127.0.0.1:7420. The paper's Programmer ⇄ Inspector design is on its legacy-open-source branch, where the map shows the two agents handing work back and forth.

CLI reference

visualagents                 guided start (same as `visualagents setup`)
visualagents demo            [--speed N]
visualagents run             [FOLDER] [--pick] [--smart] [--cwd DIR] [--no-instrument] [--no-follow-logs] [-- COMMAND…]
visualagents hook            [FOLDER] [--python PY]…
visualagents unhook | open   [FOLDER]
visualagents up              [--port 7400] [--foreground]
visualagents down | status
visualagents scan            [FOLDER] [--static | --llm] [--provider auto|anthropic|openai|google] [--model M] [--json] [-o FILE]
visualagents models          [FOLDER] [--provider anthropic|openai|google]
visualagents show | blueprint [FOLDER]
visualagents pipe            [FOLDER] [--quiet]
visualagents watch           [FOLDER] --log FILE [--from-start] [--echo]
visualagents replay          RECORDING [--project DIR] [--speed N] [--instant]

dashboard options:  --port 7420  --host 127.0.0.1  --no-browser  --blueprint FILE  --exit  --rescan

vagents is a short alias. visualagents <command> --help explains every option.

Files it creates

Privacy and security

• The dashboard binds to 127.0.0.1 only; events from your agent need a per-session token.
• Nothing leaves your machine unless you choose an LLM provider. Before code is sent, .env, key and credential files are skipped and strings that look like API keys are redacted.
• Saved keys live in ~/.visualagents/keys.json with owner-only permissions. No telemetry.
• Instrumentation never changes your agent's behaviour or output, and stays silent in processes that never call an LLM.

Troubleshooting

Development

git clone https://github.com/kaih06-hub/VisualAgents.git && cd VisualAgents
python -m venv .venv && . .venv/bin/activate
pip install -e ".[dev,all]"
pytest

The dashboard is plain HTML/CSS/JS in visualagents/web/ (no build step). New console formats go in visualagents/patterns.py, new static detectors in visualagents/static_scan.py, start-command detection in visualagents/detect.py, and runtime hooks in visualagents/_autoinstrument/visualagents_runtime.py.

License

MIT