Skill
Where it works: Claude.ai, Cowork, Claude Code, Codex, Cursor and other agents that read SKILL.md files
Best for: The guided interview, audits and written reports
TorqueSight · Process intelligence for AI agents
Your agents are a business process. Map them, run them, watch them, secure them.
TorqueSight is a skill for Claude, Claude Cowork, Claude Code and OpenAI Codex. Describe the agent system you want, or point it at one you already have. It draws a BPMN map of every agent, scores each agent's health and cost from your run logs, ranks the fixes worth making, and reviews whether your agents are safe, all in one offline viewer. It is not an agent framework. It sits above whatever runs your agents and never sees your data.
That is process intelligence for AI agents: you see inside the process your agents run.
One engine, three ways in: a skill, a CLI and an MCP server. Runs on your machine. Juggernaut never sees your data.
Works withClaude CodeClaude.aiCoworkCodexGemini CLICursorVS CodeWindsurfClaude Desktopn8n
One engine, three ways in
TorqueSight runs the same engine three ways, and all three give the same numbers.
CLI
Where it works: Any terminal
Best for: Scripts and quick runs
torquesight validate harness.yamlMCP serverNew in 1.0
Where it works: Claude Code, Codex, Gemini CLI, Cursor, VS Code, Windsurf, Claude Desktop and n8n
Best for: Agents that check pipeline health while they work, and n8n workflows that call TorqueSight as a step
The MCP server runs on your machine, uses only the Python standard library, makes no network calls and never edits your blueprints, prompts or code. Read the MCP guide
4
Surfaces
9
Console tabs
21
Platforms
0
Data sent to us
See inside the process
Logs tell you what one call did. TorqueSight shows you the whole process your agents run, and how every run moved through it.
Map it
Your blueprint becomes a BPMN map: every agent, decision, parallel step, retry loop, timer and human approval, in the lanes that own them.
Measure it
Your run logs land on the map. Each agent gets a health and a cost score, and the As run view shows which paths real runs took and how often each loop fired.
Improve it
TorqueSight ranks the fixes worth making with an estimated saving, predicts risky agents before they fail, and flags agents that could act on untrusted input without approval.
What you see that logs alone do not show
- Where work loops back, and how often it happens now compared with last month.
- Which prompt change made an agent worse, and on which date.
- Which agent takes most of the spend, and whether a cheaper model would do.
- Which path real runs take compared with the process you designed.
- Which agent reads untrusted input, touches sensitive data and can act, with no person approving first.
Built for n8n
Using n8n? Start with what you already have.
Download your workflow and import it, then import n8n's own execution history. You get the BPMN map, agent health and cost, prompt changes that made an agent worse, and ranked fixes, with no changes to your workflows.
What you get
A BPMN process map
Every agent, decision, parallel split, retry loop, timer, error handler and human approval, in a viewer you can search, zoom and share links into. Exports to BPMN 2.0 for bpmn.io and Camunda.
One viewer for everything
A single offline HTML file with nine tabs: the map, the process, health, runs, cost and fixes, security, readiness, every report, and the source blueprint and prompts.
Ranked fixes
Fifteen optimizer rules find prompt regressions, retry loops, timeouts, premium models on simple jobs and reviewers that rewrite instead of patching, each with an estimated saving and a test to run first.
A day-one risk register
Before any run, a readiness score for every agent from the blueprint and a short interview, so you know which agents are shaky before they fail.
A static security review
Each agent scored against the Rule of Two: untrusted input, sensitive data and the ability to act. Seven checks, no attacks, no code execution.
Your data stays yours
No server, no account, no telemetry. The recorder writes metadata only to your own logs, and you bring the logs to TorqueSight. Juggernaut Dev never receives anything.
How the MCP server stays safe
- The MCP server runs on your machine, over stdio or over HTTP on 127.0.0.1. It refuses any other address unless you set a token.
- It never runs a shell and keeps your API keys out of the engine's environment.
- It reads and writes only inside your project folder, never touches .env files, keys or .git, and never edits blueprints, prompts or code. A read-only mode writes no files at all.
- Every value it returns from your logs is labeled as data, because model output in logs can carry prompt injection.
What is process intelligence?
Process intelligence, also called process mining, takes the event logs a business already produces, rebuilds the real process from them, and shows where it is slow, costly or risky. Companies have used it for years on supply chains, finance and customer service.
TorqueSight brings process intelligence to AI agents. Your blueprint and your agents' run logs become a live process map, with health, cost and safety on every step.
| Aspect | Agent observability | Process intelligence with TorqueSight |
|---|---|---|
| What you look at | One trace or one model call at a time | The whole process across every run |
| The question it answers | What happened in this call? | Where does the process break, what does it cost, and what should change? |
| What you get | Traces, logs and dashboards | A BPMN map, health and cost per agent, ranked fixes, a risk register and a security review |
| People and approvals | Usually out of view | Drawn in their own lanes, with timers and escalations |
| Starting point | A tracing SDK or proxy in your code | What you already have: n8n history, OpenTelemetry traces, Langfuse exports, or a small recorder |
TorqueSight does not replace your observability tool. It reads what that tool already collects and turns it into a process you can see, score and improve.
Take the tour
The TorqueSight viewer is one HTML file that opens in any browser, offline. Pick a tab to see what it shows, then open the same tab in a live demo built from a 25-agent claims process.
Map: The process, drawn the way BPMN tools draw it.
Lanes, agents with their jobs, decisions, parallel work, retry loops, timers and human approvals. Switch between Designed and As run, color the steps by health, spend or latency, search for a step, and move around a long process with the minimap, the keyboard and full screen. Clicking a step opens it and puts a link to that step in the address bar, so you can send it to a colleague.
Open the Map tab in the live demo
Process: One card per step, in run order.
Each card shows the step's job, what it receives and from whom, the context it is given, what it produces, where it hands off, its timers and error handlers, its model and limits, and its readiness and security scores.
Open the Process tab in the live demo
Health: Which agents are slipping.
Every AI agent is a bubble placed by health score and sized by spend, with a watchlist of drift, prompt versions and budget runway. Click an agent for its 30-day trajectory and exactly what is pulling its score down.
Open the Health tab in the live demo
Runs: What happened in each run.
Every recent run by score and cost, the runs that failed or went over budget, and a step-by-step timeline of any run.
Open the Runs tab in the live demo
Optimize: Where the money goes and what to fix first.
Spend by agent over 30 days, ranked fixes with their evidence and estimated savings, and a what-if table to try a cheaper model before you change anything.
Open the Optimize tab in the live demo
Security: Are the agents safe?
Each agent's exposure to three risk properties: untrusted input, sensitive data and the ability to act. Agents with all three and no human approval rise to the top, each with its fix. Static review only; nothing is attacked.
Open the Security tab in the live demo
Readiness: The day-one risk register.
Which agents are likely to struggle, predicted from the design and the interview before any run, so you can fix them before they fail.
Open the Readiness tab in the live demo
Reports: Every report, readable in place.
The process spec, the optimize report, the risk register and the security register, rendered and ready to download.
Open the Reports tab in the live demo
Source: The blueprint and every prompt.
The whole blueprint and each prompt file, read-only, with copy and download, so the console is the one place to see everything.
Open the Source tab in the live demoWhat is BPMN, and why it matters
BPMN (Business Process Model and Notation) is the international standard for drawing a process, published as ISO/IEC 19510. Its shapes are simple and consistent, so a developer, a business owner and an auditor can read the same picture. Here is a small agent process drawn the way TorqueSight draws yours.
- Start eventWhat kicks the process off: an email, a form, a schedule.
- AI agent taskOne agent with one job. TorqueSight scores each of these.
- DecisionA branch, such as approved or not, covered or not.
- Retry loopWork sent back to an earlier agent. How often this happens is a key health signal.
- LaneWho owns the steps: a team, a role, or AI agents versus people.
- Human approvalWhere a person must decide before the process continues.
- TimerWhat happens when something takes too long, such as a warning after 24 hours.
- End eventHow the process finishes.
Everyone reads the same picture
Code shows one developer what happens. A BPMN map shows the whole team, including the people who approve, audit or pay for the work.
It shows what code hides
Where agents hand off, where people approve, where work loops back, where a timeout fires, and where a mistake can reach a customer.
It is a standard
The same map exports as BPMN 2.0 and opens in bpmn.io or Camunda, so it fits the tools process teams already use.
In TorqueSight, the map is live
Steps are colored by health, spend or latency, and As run shows which paths real runs took and how often each loop fired.
Works with what you already use
Bring your own logs. TorqueSight reads a blueprint and a run log and never connects to your systems. Pick your stack to see exactly how it connects.
Import what you already have
No code changes. TorqueSight reads history your platform already keeps.
Frameworks that speak OpenTelemetry
Send their traces to a collector you run, then import the file.
Add the recorder
One file with no dependencies, called from the place every model call passes through.
20 platform guides, each also available as its own page.
Open a live demo
Each demo is the real console, a single offline HTML file built from sample data. It opens in a new tab.
Claims settlement
A 25-agent claims process with 10 human approvals. The flagship example: the BPMN map, health, cost, ranked fixes, the security review and every report.
n8n support triage
Built entirely from an n8n workflow export and its execution history, with no hand-written blueprint. It finds a drafter that got worse after its system message was edited.
All five pipelines
One console across claims, support triage, research, marketing content and code review, to see how the viewer handles several pipelines at once.
Questions
What is TorqueSight?
An open-source skill and toolkit that brings process intelligence to AI agents. It maps your agents as a BPMN process, scores each agent's health and cost from your run logs, ranks the fixes worth making, and reviews whether the agents are safe.
Is TorqueSight an agent framework or a harness?
No. It does not run your agents. It sits above whatever already runs them, such as n8n, LangGraph, CrewAI, the OpenAI Agents SDK or the Claude Agent SDK, and reads the history they produce.
How is it different from agent observability tools?
Observability tools show what happened inside a single call or trace. TorqueSight shows the whole process across every run: where it breaks, what it costs and what to fix. It can import data from tools such as Langfuse and OpenTelemetry.
Does TorqueSight see my data?
No. There is no TorqueSight server, account or telemetry. It runs in your Claude or Codex session and on your machine, reads logs you provide, and its recorder writes metadata only to your own logs.
Which platforms does it work with?
n8n, OpenTelemetry, Langfuse, LangChain, LangGraph, CrewAI, the OpenAI Agents SDK, the Claude Agent SDK, Google ADK, Strands, Mastra, the Vercel AI SDK, Pydantic AI, LlamaIndex, Amazon Bedrock, Bedrock AgentCore, Azure AI Foundry, Lovable, the Anthropic and OpenAI APIs, and xAI Grok.
Is it free?
Yes. TorqueSight is open source under the Apache-2.0 license. TorqueSight Enterprise, a delivery kit deployed in your own cloud, is coming soon.
Do I need the MCP server if I use Claude.ai?
No. Claude.ai and Cowork use the skill, which runs the same engine inside Claude's sandbox. The MCP server is for apps that run on your computer, such as Claude Code, Codex, Gemini CLI, Cursor, VS Code, Windsurf, Claude Desktop and n8n.
Is the MCP server safe to run next to my agents?
It runs locally, makes no network calls, never runs a shell, and reads and writes only inside your project folder. It never edits your blueprints, prompts or code, and it labels everything from your logs as data so an injected instruction in a log is never followed. A read-only mode writes no files at all.
Does TorqueSight send my data anywhere?
No. The skill, CLI and MCP server all run where you run them. Juggernaut never sees your blueprints, prompts or run logs.
Which MCP versions does it support?
Both protocol eras: the initialize handshake from 2024-11-05 through 2025-11-25, and the stateless 2026-07-28 revision. It is tested against the official MCP Python SDK over stdio and HTTP.
Can n8n Cloud use the MCP server?
n8n Cloud cannot reach a program on your computer. Use the MCP server with self-hosted n8n, or export your execution history and run TorqueSight on it with the skill or the CLI.
Install
The fastest start, for Claude Code, Codex, Cursor and most other coding agents:
npx skills add YOURORG/torquesightThat installs the skill. Pick your app below to add the MCP server's typed tools as well.
Install the plugin. It brings the skill, the /torquesight:audit, /torquesight:health and /torquesight:optimize commands, the reviewer agent and the MCP server, which starts on its own.
/plugin marketplace add YOURORG/torquesight
/plugin install torquesight@juggernautRestart Claude Code, then run /mcp to confirm torquesight is connected.
- Turn on code execution in Settings.
- Download torquesight.skill from the latest release.
- Go to Settings, Capabilities, Skills, choose Upload skill and upload torquesight.skill.
- Start a new chat and ask: Use TorqueSight to audit this n8n workflow.
Claude.ai on the web cannot reach a program on your computer, so it uses the skill, which runs the same engine inside Claude's sandbox.
npx skills add YOURORG/torquesight
codex mcp add torquesight -- python3 /path/to/torquesight/mcp-server/torquesight_mcp.pyConsole builds on large logs can take longer than Codex's default tool timeout, so raise it in ~/.codex/config.toml:
[mcp_servers.torquesight]
command = "python3"
args = ["/path/to/torquesight/mcp-server/torquesight_mcp.py"]
tool_timeout_sec = 300gemini extensions install https://github.com/YOURORG/torquesightThe extension bundles the MCP server and a context file that teaches Gemini the TorqueSight workflow.
Clone the repository, then print the exact setup with every path filled in:
git clone https://github.com/YOURORG/torquesight.git
cd torquesight
python3 mcp-server/torquesight_mcp.py wrap cursorUse wrap windsurf for Windsurf. For Cursor, the result goes in ~/.cursor/mcp.json and looks like this:
{
"mcpServers": {
"torquesight": {
"command": "python3",
"args": ["/path/to/torquesight/mcp-server/torquesight_mcp.py"]
}
}
}Add this to .vscode/mcp.json in your project:
{
"servers": {
"torquesight": {
"type": "stdio",
"command": "python3",
"args": ["/path/to/torquesight/mcp-server/torquesight_mcp.py"]
}
}
}Open Settings, then Developer, then Edit Config, add this block, and restart Claude Desktop. The --root folder is the project TorqueSight may read.
{
"mcpServers": {
"torquesight": {
"command": "python3",
"args": ["/path/to/torquesight/mcp-server/torquesight_mcp.py", "--root", "/path/to/your/project"]
}
}
}- Start the server in the folder that holds your blueprints and run logs:
python3 /path/to/torquesight/mcp-server/torquesight_mcp.py --http --port 8765 - In n8n, add an MCP Client Tool node to your AI Agent.
- Set Server Transport to HTTP Streamable and Endpoint to http://127.0.0.1:8765/mcp. If n8n runs in Docker Desktop on the same machine, use http://host.docker.internal:8765/mcp.
n8n in Docker on Linux, or on another machine, needs a token and Header Auth. The MCP guide covers it.
Check your setup
python3 mcp-server/torquesight_mcp.py doctorIt checks Python, PyYAML and the engine, and tells you what to fix.
Requirements: Python 3.8 or newer, and PyYAML for YAML blueprints. Nothing else.
Why an MCP server
The skill teaches an agent how to interview you and read the results. The MCP server lets any agent ask TorqueSight questions while it works.
Agents check their own pipeline
An agent can ask which stage in its last 50 runs has the worst retry rate and get structured rows back, instead of reading log files.
n8n calls it as a step
The MCP Client Tool node connects to TorqueSight, so a nightly workflow can run a health check after each batch and branch on the result.
Works beyond Claude
Codex, Gemini CLI, Cursor, VS Code and Windsurf get the same tools and the same numbers, with no new skill format.
Typed tools, not flags
Every tool has a schema, pagination and evidence, and the audit, health, optimize and security prompts show up as commands in clients that support them.
11 tools
torquesight_validate_blueprintSchema and rule errors, and unmapped steps in the run logs
Read onlytorquesight_get_process_mapThe designed process as steps and connections
Read onlytorquesight_importBlueprints from n8n or BPMN; run logs from n8n executions, OpenTelemetry or Langfuse
torquesight_get_stage_healthPer-agent success, retries, latency, tokens and score
Read onlytorquesight_find_bottlenecksThe worst failure points, with the run ids that prove them
Read onlytorquesight_compare_windowsBefore and after a prompt, model or workflow change
Read onlytorquesight_get_runOne run's path, failures and loop decisions
Read onlytorquesight_optimizeRanked fixes with evidence and a proposed patch that is never applied
torquesight_security_reviewA static review of the agents themselves that never generates attacks
torquesight_risk_registerDesign readiness before any logs exist
torquesight_build_outputsThe offline console, BPMN 2.0, SVG map and Markdown spec
Built for n8n
Import your workflow and execution history, and every AI agent gets a score. Then connect the MCP Client Tool node so your own n8n agent can ask TorqueSight which step is failing, with no changes to the workflow it is watching.
python3 /path/to/torquesight/mcp-server/torquesight_mcp.py --http --port 8765Endpoint http://127.0.0.1:8765/mcp, transport HTTP Streamable.
See the n8n tab in InstallWhere it is headed
In 1.0
- The skill, with quick and deep interviews
- The Claude Code plugin, with commands and the reviewer agent
- The CLI
- The MCP server for Claude Code, Codex, Gemini CLI, Cursor, VS Code, Windsurf, Claude Desktop and n8n
- Importers for n8n, OpenTelemetry, Langfuse and BPMN
- Agent health scoring, the optimizer and the static agent security review
- Design readiness before any logs exist
- The offline console and BPMN 2.0 export
Next
- Security findings mapped to NIST control families
- MCP tool scoring: every MCP server your agents use becomes a scored step, including the context spent loading tool definitions
- A Docker sandbox that validates security findings with harmless probes only
- A live local viewer that follows runs as they happen
Available now
TorqueSight, open source
Design, audit, health, optimize, the risk register, the security preview, the nine-tab viewer, and importers for n8n, OpenTelemetry and Langfuse. Bring your own logs.
Coming soon
TorqueSight Enterprise
A delivery kit deployed in your own cloud that copies logs from CloudWatch, Azure Monitor, Datadog or Splunk on a schedule and hosts the console behind your own sign-in. Your data never leaves your account.
In the future
TorqueSight Cloud
Hosted dashboards, alerts and history, plus the full visualization and security experience in one place. Opt-in, metadata only.
Get it on GitHub →Take the tourRead the example use casesWork with Juggernaut DevMore from Juggernaut Dev open source
TorqueSight is open source under Apache-2.0, built by Juggernaut Dev. Works in Claude.ai, Claude Cowork, Claude Code and OpenAI Codex.