Agent-agnostic positioning: MCP-first, Claude Code as the verified path

.claude-plugin/marketplace.json

@@ -1,12 +1,15 @@
 {
   "name": "stackcanvas",
-  "description": "Live infrastructure canvas for Claude Code.",
-  "owner": { "name": "kp", "url": "https://github.com/pshenok" },
+  "description": "Live infrastructure canvas for AI coding agents.",
+  "owner": {
+    "name": "kp",
+    "url": "https://github.com/pshenok"
+  },
   "plugins": [
     {
       "name": "stackcanvas",
       "source": "./plugin",
-      "description": "Live infrastructure canvas for Claude Code: watch the agent build your Terraform, drag new resources, the agent writes the HCL.",
+      "description": "Live infrastructure canvas: watch your agent build Terraform, drag new resources, the agent writes the HCL. This plugin wires it into Claude Code.",
       "category": "infrastructure"
     }
   ]

README.md

@@ -1,6 +1,6 @@
 # stackcanvas
 
-Live infrastructure canvas for [Claude Code](https://claude.com/claude-code).
+Live infrastructure canvas for AI coding agents — any agent that speaks [MCP](https://modelcontextprotocol.io) ([Claude Code](https://claude.com/claude-code) is the CI-verified path).
 The agent writes and plans your Terraform — stackcanvas shows it as a living
 diagram. Drag new resources onto the canvas; the agent turns them into
 idiomatic HCL. No SaaS, no credentials leave your machine: everything runs on
@@ -23,7 +23,13 @@ localhost, reading your local state and plan.
    canvas highlights what will change. Only the agent executes Terraform —
    the canvas has no apply button by design.
 
-## Install (Claude Code)
+## Install
+
+stackcanvas is a standard MCP stdio server (`npx -y stackcanvas`) — point any
+MCP-capable coding agent at it. The paths below, in order of how battle-tested
+they are:
+
+### Claude Code (CI-verified)
 
     claude plugin marketplace add pshenok/stackcanvas
     claude plugin install stackcanvas@stackcanvas
@@ -42,12 +48,12 @@ Or without the plugin system:
 Then, inside a repo with Terraform, just ask: *open the stackcanvas canvas
 for this repo*.
 
-## Other MCP clients
+### Other MCP clients
 
-stackcanvas is a standard [MCP](https://modelcontextprotocol.io) server, so
-any MCP-compatible client can run it. The snippets below are **reported to
-work as standard MCP servers; not yet CI-verified** — only the Claude Code
-path above is exercised in CI.
+The snippets below are **reported to work; not yet CI-verified** — only the
+Claude Code path above is exercised in CI. Codex CLI and other MCP-capable
+agents should work with the equivalent stdio config (`npx -y stackcanvas`) —
+untested, reports welcome in issues.
 
 **Cursor** (`.cursor/mcp.json`):
 

docs/PROBLEM.md

@@ -60,7 +60,7 @@ Supervision fail-safety: if the state is corrupt or terraform fails, the canvas
 1. The palette (provider packs — AWS first, plus an "any terraform type" field for everything else) → drag → a **dashed draft node** appears on the canvas. A draft is not reality — the visual grammar is strict: what exists is solid, what's intended is dashed.
 2. You pull edges from the draft to existing nodes ("connect it to *this* VPC" — as a gesture, not an address from memory). The inspector has a free-text "wishes" field ("db.t4g.micro, no multi-AZ"). Right-click on an existing node — "request change" or "mark for removal" (rendered struck-through, dashed). An edge drawn between two existing nodes = a "connect them" request.
 3. The **Send to agent** button collects everything into a single intent JSON with strict resource addresses: `{add: [{type, wishes, connect_to: ["aws_vpc.main"]}], modify: [{address, wishes}], remove: [{address}]}`. No "which of the three webs" ambiguity — addresses come from the graph, not from the human's memory.
-4. The agent receives the intent through a blocking MCP tool, `await_canvas_intent`: it waits in a loop (45-second iterations — tuned to MCP client request timeouts, a number discovered by live testing), and the moment you hit Send to agent, the intent arrives as the tool's return value. A bundled skill teaches Claude Code the whole cycle: open the canvas → run a plan → await intents → write HCL → plan again → await again. The user runs one command and then lives in the browser with a mouse.
+4. The agent receives the intent through a blocking MCP tool, `await_canvas_intent`: it waits in a loop (45-second iterations — tuned to MCP client request timeouts, a number discovered by live testing), and the moment you hit Send to agent, the intent arrives as the tool's return value. A bundled skill teaches the agent the whole cycle (shipped for Claude Code; the tool descriptions alone carry the loop for other MCP clients): open the canvas → run a plan → await intents → write HCL → plan again → await again. The user runs one command and then lives in the browser with a mouse.
 5. **Only the agent writes HCL.** The tool contains zero code-generation templates — a deliberate architectural bet. Template codegen (the Brainboard path) is an eternal race to cover thousands of resource types, producing code that's foreign to your repo. The agent writes in *your repository's* style, knows its modules and conventions, and covers all of Terraform at once. Our "codegen" gets cheaper and smarter with every model release — for free.
 6. The loop closes by itself: the agent writes code → runs a plan → the watcher notices → drafts dissolve into real nodes with plan statuses → the human reviews the diff → tells the agent to apply. If no agent is waiting (or you use a different client), a "Copy as prompt" button turns the intent into a ready-to-paste text prompt.
 
@@ -75,11 +75,11 @@ The concrete mechanisms that make trusting us unnecessary:
 - **The server binds to `127.0.0.1` only.** Not "behind a login" — physically unreachable from outside. There is no cloud of ours to run your infrastructure through — state, HCL, and credentials have nowhere to go. (One honest exception: an **opt-in** anonymous telemetry counter — install/reopen/intent counts only, never resource data — declinable at first run, documented in [TELEMETRY.md](../TELEMETRY.md).)
 - **The tool holds no credentials at all** (v1): it reads local state/plan files. Even the future account scan uses your local `~/.aws` profile with read-only sessions.
 - **Secrets are masked before they reach the UI:** everything the state marks `sensitive` (passwords, keys) becomes `•••` inside the parser — verified by live testing against `random_password`.
-- **The agent is your own.** We don't proxy tokens and don't embed our own LLM: it's the Claude Code (or any MCP client) you already pay for, with your key.
+- **The agent is your own.** We don't proxy tokens and don't embed our own LLM: it's whatever MCP-capable coding agent you already pay for — Claude Code, Cursor, Windsurf — with your key.
 - **The UI has no `terraform apply` control — by design.** The primary button is "Send to agent": it hands your intent over, nothing more. The canvas proposes; only the agent executes, and the skill explicitly forbids `terraform apply` without an explicit human request. Today this is supervision *by convention*; in the Studio stage it becomes supervision *by construction* — a credential broker will physically withhold mutate-scope credentials until a human approves.
 
 ### Packaging: why it spreads on its own
 
-Everything ships as one npm package (`npx stackcanvas` = an MCP server with the UI embedded) plus a Claude Code plugin (config + skill). Installation is one command; Cursor and Windsurf are reported to work as standard MCP clients; Claude Code is the CI-verified path. Extension is data, not code: a new cloud = a PR with a provider pack (palette + containment rules + icons) — a deliberate moat mechanic: the gesture→HCL corpus and provider packs accrete in the OSS repo the way oh-my-zsh accretes plugins.
+Everything ships as one npm package (`npx stackcanvas` = a standard MCP stdio server with the UI embedded) plus a Claude Code plugin (config + skill) for one-command setup there. Installation is one command; Cursor and Windsurf are reported to work as standard MCP clients; Claude Code is the CI-verified path. Extension is data, not code: a new cloud = a PR with a provider pack (palette + containment rules + icons) — a deliberate moat mechanic: the gesture→HCL corpus and provider packs accrete in the OSS repo the way oh-my-zsh accretes plugins.
 
 **The anti-scope is part of the solution:** not a PaaS (we don't hide Terraform), not monitoring (we won't rebuild Datadog — only signals available with zero setup), never SaaS, no codegen of our own. Every "not" keeps the product in the one empty niche: **eyes and a steering wheel for the human whose agent already has hands.**

packages/mcp/package.json

@@ -1,6 +1,6 @@
 {
   "name": "stackcanvas",
-  "version": "0.2.2",
+  "version": "0.2.3",
   "type": "module",
   "main": "dist/cli.js",
   "bin": {
@@ -29,7 +29,7 @@
     "@types/node": "^22.0.0",
     "tsup": "^8.0.0"
   },
-  "description": "Live infrastructure canvas for Claude Code: see what your AI agent is about to do to your Terraform \u2014 and approve it \u2014 before it happens. Local-first, credentials never leave your machine.",
+  "description": "Live infrastructure canvas for AI coding agents (MCP): see what your agent is about to do to your Terraform \u2014 and approve it \u2014 before it happens. Local-first, credentials never leave your machine.",
   "license": "MIT",
   "repository": {
     "type": "git",

packages/mcp/src/version.ts

@@ -10,4 +10,4 @@
 // script) is release-engineering scope (docs/SPEC.md Telemetry chapter, §5 /
 // increment 8), not part of the emitter wiring here — keep the two in sync
 // by hand until that lands.
-export const VERSION = '0.2.2'
+export const VERSION = '0.2.3'

plugin/.claude-plugin/plugin.json

@@ -1,10 +1,20 @@
 {
   "name": "stackcanvas",
-  "description": "Live infrastructure canvas for Claude Code: watch the agent build your Terraform, drag new resources, the agent writes the HCL.",
-  "version": "0.2.2",
-  "author": { "name": "kp", "url": "https://github.com/pshenok" },
+  "description": "Live infrastructure canvas: watch your agent build Terraform, drag new resources, the agent writes the HCL. This plugin wires it into Claude Code.",
+  "version": "0.2.3",
+  "author": {
+    "name": "kp",
+    "url": "https://github.com/pshenok"
+  },
   "homepage": "https://github.com/pshenok/stackcanvas",
   "repository": "https://github.com/pshenok/stackcanvas",
   "license": "MIT",
-  "keywords": ["terraform", "opentofu", "infrastructure", "canvas", "diagram", "mcp"]
+  "keywords": [
+    "terraform",
+    "opentofu",
+    "infrastructure",
+    "canvas",
+    "diagram",
+    "mcp"
+  ]
 }

plugin/.mcp.json

@@ -1,5 +1,5 @@
 {
   "mcpServers": {
-    "stackcanvas": { "command": "npx", "args": ["-y", "stackcanvas@0.2.2"] }
+    "stackcanvas": { "command": "npx", "args": ["-y", "stackcanvas@0.2.3"] }
   }
 }