The Short Version
Sixty seconds, from install to a cloud your AI understands.
Sound on. Everything below is the same story in text, in case you would rather read it.
OpsCanvas MCP · Free for individual engineers
The Cloud Context Your AI Needs
Use AI to manage your cloud.
A free, local MCP server for the Claude and OpenAI tools you already use. It maps your whole cloud on your own machine in minutes.
41Seconds to a live map of your cloud
35/35Answers correct. 0/35 without it
LocalRead only, nothing uploaded
$0Marginal cost on your AI subscription
Install
npm install -g @opscanvas/opscanvasFree with an OpsCanvas account. Works today with Claude Code (CLI), Claude Desktop, Claude Cowork, OpenAI Codex, and ChatGPT Desktop.
What is OpsCanvas MCP?
OpsCanvas MCP is a free MCP server that runs locally on your own workstation and gives the AI tool you already use, Claude Code (CLI), Claude Desktop, Claude Cowork, OpenAI Codex, or ChatGPT Desktop, a live and correlated map of your cloud. That map is the Cloud Intelligence Graph™. It is built for the individual engineer: it uses the credentials already on your machine, it is read only, and nothing about your estate is uploaded or hosted anywhere. Team capabilities, with one shared hosted map, are a separate paid edition.
Getting Started
Running in about two minutes.
No agent to adopt, no tagging project, no data to hand over. You stay in the tool you already work in.
Why Now
Finding problems got cheap. Fixing them did not.
AI now ships software faster than anyone can manage what it builds, and the tools that watch it are better at finding than at finishing.
- About $250 billion of cloud spend is wasted every year, and the waste rate just rose for the first time in five years
- The tools find well, and some will draft a fix, but someone still has to carry it out. Under 1% of findings ever get fixed
- Most waste and risk is already known and deliberately left alone, because nobody can say quickly what depends on it
- So the question that stops every change is the same one, every time: what breaks if I change this?
The Missing Piece
Your AI sees your cloud in pieces, so it guesses at how they connect.
A model can read an API response. What it cannot do is correlate six accounts, two clouds, five Kubernetes contexts, your pipelines, and your bill into one picture of what depends on what. That correlated context is private, tightly held, and exists in no training corpus. The Cloud Intelligence Graph™ is that map, wired by real dependency rather than by tags that went stale on day two.
For the Skeptics
The same model, measured with the Cloud Intelligence Graph™ and without it.
Thirty-five complex questions about a real cloud estate, put to the same LLM twice.
35/35
Correct with the graph
Every environment question answered correctly, with the evidence trail behind each answer.
0/35
Correct without it
Without the graph the model mostly declined to answer rather than fabricate, which is the right call and still leaves you with nothing.
$0
Marginal cost
Bring your own model. OpsCanvas works through the AI subscription your team already pays for, with no separate per-token bill.
The Question That Stops Everything
"What breaks if I change this?"
Dependencies come back with the answer. Ask whether it is safe to fail over a database tonight and you get the blast radius alongside the answer: what sits downstream, what has no retry, and what is genuinely untouched. That is the precondition for every safe change that follows, and it is the difference between leaving something alone and fixing it.
Where The Data Goes
The map is built on your machine, under your own credentials.
It reads what this workstation can already reach, and you confirm every source before it enters your map. Your estate is not uploaded, not used for training, and not shared. The model sees answers. It does not get your infrastructure.
Accurate, Secure, Accountable
The map never leaves. The answers do.
Three questions decide whether you point an AI at your cloud: will it be right, will my data leave, and who is answerable when it acts. OpsCanvas answers all three structurally rather than by policy. It is never another agent. It is the grounding and the guardrails for the AI you already trust.
- Runs locally, your cloud data stays on your machine
- Read-only by default
- Limited to your own permissions, never beyond your access
- Your credentials never leave this machine, and you confirm every source
- A human approves every change, step by step
- Full audit trail with decision traces
Inside Your LLM
Ask your cloud anything, in plain English.
Ask from Claude Code (CLI), Claude Desktop, Claude Cowork, OpenAI Codex, or ChatGPT Desktop. Every answer is grounded in your live topology, with the evidence trail behind it.
your LLM
>
1,180 resources across 6 accounts, sized by monthly spend and wired by real dependency.
Biggest single line is ec2-ml-training at $28.6k a month, with no CPU activity in 38 days.
✓ Spend, waste, health, and ownership are already on the map.
This is a configuration change, not organic growth. Spend on the payments path jumped 31% at 02:14.
The update the Payments team deployed overnight is responsible. It doubled the read replica count on rds-payments.
✓ Correlated against deploy history, not inferred from the bill alone.
Six resources sit downstream, including the API fleet and the CDN origin.
Your analytics estate is untouched, so a failover does not affect reporting.
⚠ Two of the six have no retry configured. Review before you proceed.
Cross-checked cost against topology and connection history.
4 zombies: a load balancer with 0 connections for 6 months, 2 unattached volumes, 1 idle NAT gateway.
$212/mo recoverable. None have downstream dependents.
Answers grounded in your Cloud Intelligence Graph™ · illustrative output
Illustrative session from a sample environment.
Why It Compounds
You fix things, and the same map shows the difference.
Every question, investigation, and change leaves something behind: your operational memory, and the proof of what actually worked. The map is your own context and history of your own cloud, and it gets stronger every time it is used. Models keep getting better, and none of that improvement hands an LLM the correlated context of your cloud, because that context is private and tightly held.
Works With Your AI
It plugs into the tools your engineers already use.
OpsCanvas arrives as an MCP server, so there is no new agent to adopt and no new interface to learn. Same tool, same subscription, now with your live cloud topology behind every answer.
Five surfaces are tested and supported today, marked Supported below. More are close behind: because this is a standard MCP server, it will often work in a tool we have not listed yet, but we only call something supported once we have tested it and done the tuning that tool needs.
Supported
Supported
Supported
Supported
Supported
Coming soon
Coming soon
Coming soon
Coming soon
Coming soon
Common Questions
OpsCanvas MCP, answered.
What is MCP?
The Model Context Protocol is an open standard that lets AI tools like Claude Code (CLI), Claude Desktop, Claude Cowork, OpenAI Codex, and ChatGPT Desktop call external capabilities. OpsCanvas arrives as an MCP server, which is why it works inside the AI your engineers already use instead of asking them to adopt another agent.
Which AI tools does it work with?
Tested and supported today: Claude Code (CLI), Claude Desktop, Claude Cowork, OpenAI Codex, and ChatGPT Desktop. Cursor, OpenCode, Antigravity CLI, Gemini Code Assist, and Grok are coming soon. OpsCanvas MCP is a standard MCP server, so it will often work in an MCP-capable tool we have not listed. Coming soon means we have not finished testing it or done the tuning that tool needs, not that it is blocked.
Does my cloud data leave my machine?
No. The map is built locally, under your own credentials, and it stays on your machine. Your estate is not uploaded, not used for training, and not shared. The model sees answers. It does not get your infrastructure.
What can I do in the first session?
Read and understand your estate. Ask for the topology of your most active application, what is costing you the most right now, what nothing is using, what changed overnight, and what sits downstream of anything you are thinking of touching.
How long does the map take to build?
Minutes. In the walkthrough above, a six-account estate of 1,180 resources reached a live map in 41 seconds. Larger estates take longer, and you confirm every source before it enters the map.
Do I need to tag my resources first?
No. The map is built from runtime behavior and real dependency, not from tags. Tagging projects are incomplete on day one and wrong by day two, which is exactly the gap this closes.
Can it change things in my cloud?
Not in this release. OpsCanvas MCP is read-only today by design, on the principle that trust comes before mutation. It reads, correlates, and explains your estate so you can decide what to change and make the change yourself. Guardrailed remediation with human approval on every write is in development and is not part of what you install today.
What does it cost?
Nothing. OpsCanvas MCP is free with an OpsCanvas account, and running it costs $0 extra because it works through the AI subscription you already pay for rather than billing you per token. There is no credit card and no trial clock.
Is this the local version or the hosted one?
Local. This is the free edition for the individual engineer: the server runs on your workstation, the map is built and stored there, and there is nothing to provision and no tenant to sign up for. The hosted edition, where a team shares one map, is a separate paid product and is not what the install command above gives you.
Do I need to install anything else, or get my cloud team involved?
No. It reads what your workstation can already reach, using the cloud credentials you have already configured, so there is no new role to provision, no agent to deploy into your accounts, and nothing for anyone to approve before you try it. You confirm each source before it enters your map.
Why not just connect my assistant directly to my cloud accounts?
Direct connections give the model raw data, and in our evaluation that was not enough to produce a single correct environment answer: 35 of 35 correct with the graph connected, zero without. Dependencies, blast radius, ownership, and change history live in the graph, not in any single API.
How is this different from the Oscar CLI?
Oscar is our own command line agent and it continues to ship and work. OpsCanvas MCP is the same graph and the same safety model, delivered inside the tools your engineers already use rather than as another agent alongside them.
Get Started
Install once, then ask your model anything about your cloud.
Free with an OpsCanvas account. Read only, local, and $0 marginal cost on the subscription you already have.