Pmx Canvas

PMX Canvas

PMX Canvas is a server-authoritative spatial workbench controlled through MCP, HTTP, or the CLI. Humans curate agent context by pinning nodes; agents read that curation through canvas://pinned-context. State survives browser refresh.

Runtime prerequisites

PMX Canvas 0.6.1 requires Bun >=1.4.2, including when installed through npm. Check bun --version and ensure Bun is on the MCP host's PATH, not just your shell's; use an absolute executable path if needed. Install/configure only when requested. For pinned installation, MCP configuration, managed services, and disposable verification, read Installing PMX Canvas.

Required Operating Sequence

0. After a pmx-canvas upgrade, refresh this skill first. Run pmx-canvas skills sync --yes (--check to detect drift; exit 1 when stale) — it finds the skill copies already installed in the workspace (whatever layout your agent uses) and replaces the complete trees from the package, references/evals/fixtures included. Sync ONLY from the installed package, never from a source checkout: this skill documents the runtime it ships with, and a newer repo skill paired with an older installed runtime advertises commands and themes the binary does not have (0.4.3 report skew finding). Also restart the HTTP daemon and any --mcp processes; Bun does not hot-reload a running process.
1. Open or focus the workbench before mutating. Reuse one visible canvas surface for the session.
2. Verify workspace identity at the actual target. Read GET /health at the URL your CLI/MCP/browser is using; its workspace must equal the intended canonical absolute root. pmx-canvas serve status --port=<port> checks the local listener on that port, not an arbitrary PMX_CANVAS_URL. A healthy listener may belong to another project.
3. Read before write. Search with canvas_query { action: "search", query } before creating nodes. Read the full layout only when necessary. The MCP parameter is query — passing the HTTP API's q is silently ignored and returns zero results.
4. Snapshot before destructive changes. Use canvas_snapshot { action: "save", name } before clear, restore, or a major reorganization.
5. Show intent with the Ghost Cursor — by default. Signal with canvas_intent { action: "signal", ... } before every meaningful create, move, connect, remove, or edit, then pass the returned intent.id as intentId on the mutation so the ghost settles into the result. Use it as much as possible to make your next move and your work visible: the human watches intent form and can veto mid-thought. Skip it only for trivial in-place tweaks or high-frequency batch churn. The default TTL (~8s) expires between agent turns: signal with ttlMs: 30000 and settle by passing intentId on the mutation in the same or next call. Since 0.4.5, an agent mutation WITHOUT a signal still shows a server-synthesized auto-ghost (rendered lighter — dimmer, dotted, no veto — and settled instantly). That is a visibility floor, not a replacement: only an explicit signal gives the human a real pre-mutation veto window, your reasoning (reason), and staged multi-step previews. Batch and browser-human actions never auto-ghost.
6. Attach a session so the human can see you (0.4.8+). Start board work with canvas_ax_state { action: "set-presence", attached: true, label: "<who you are>" } and end it with { attached: false } — detaching hands the human a receipt. Everything the session gives you and asks of you (cursor + phase chip, the session panel, steering, the scope fence's 403s, the human edit lock's 409s, unattended approvals) is in Sessions & the human below.
7. Mutate through current composites. Prefer the 16 composite MCP tools below.
8. Arrange and validate. After batch changes, use canvas_view { action: "arrange" } when appropriate and always finish with canvas_query { action: "validate" }. 7b. Show the human. After creating user-facing output, bring the camera to it: a single node gets canvas_view { action: "focus", id } (pans by default), a small cluster gets canvas_view { action: "fit", nodeIds: [...] } with exactly the new ids. See In-View Placement & Sizing below — auto-placement is board-relative, not camera-relative.
9. Verify context pins. Pin with canvas_pin_nodes or the browser's Pin as context, then read canvas://pinned-context.
10. Clean up temporary nodes. Remove retry/test fixtures and restore the baseline snapshot when the task requires leaving the board unchanged.

Workspace Safety

Before any create, update, remove, clear, restore, arrange, or smoke test, choose an absolute project root and a dedicated free port. For a local target:

export PMX_CANVAS_WORKSPACE_ROOT=/absolute/path/to/project
export PMX_CANVAS_PORT=14313 # Example only: choose a free port for this workspace.
unset PMX_CANVAS_URL # Remove a stale URL before selecting a different local target.
curl --fail --silent --show-error "http://localhost:${PMX_CANVAS_PORT}/health"
pmx-canvas serve status --port="$PMX_CANVAS_PORT"

Compare the health field workspace (not workspaceRoot) to the intended root, resolving symlinks before comparison. Stop on mismatch or failed health; do not write. If deliberately using PMX_CANVAS_URL or --server-url, check /health at that exact URL instead. Confirm the browser and MCP use the same verified target, especially after fallback to another port. Do not print URLs containing credentials.

pmx-canvas smoke reports health/workspace/version information, checks MCP initialize, creates/searches/removes a temporary node, and validates the board. It mutates state and does not enforce your intended workspace identity; version skew is reported, not necessarily a failing check. Use a disposable workspace, verify health first, and inspect both the JSON details and exit status (1 when a check fails). It is not a read-only probe or a guarantee of unchanged history.

• If responsive: true but pidRunning: false, treat the listener as potentially stale.
• On mismatch, leave the other workspace's server alone. Start the intended workspace on a different free port, target it explicitly, and re-check /health before writing.
• On an ordinary local machine, pmx-canvas serve --daemon --no-open --port=<free-port> refuses a different-workspace occupant. In a managed environment, supervise foreground pmx-canvas serve --no-open --port=<free-port> with the host's service manager instead.
• PMX_CANVAS_PORT selects the CLI/MCP port. CLI target flags override environment values (--server-url wins over --port); otherwise PMX_CANVAS_URL wins over PMX_CANVAS_PORT. For server startup, explicit --port wins, then PMX_WEB_CANVAS_PORT, then PMX_CANVAS_PORT; in Amp orbs PORT is a later fallback. URL targeting does not select the server's bind port.
• MCP transport workspace resolution. Set PMX_CANVAS_WORKSPACE_ROOT plus a dedicated PMX_CANVAS_PORT in the host configuration. Without an explicit root, MCP may inherit a different workspace from the preferred-port daemon; PMX_CANVAS_ALLOW_WORKSPACE_SPLIT=1 opts out of that heuristic. Current source honors an explicit root, but may bind a fallback port if the preferred port is occupied. Root alone therefore does not guarantee CLI, MCP, and browser target alignment; older installed runtimes can also differ. Inspect startup output and the actual health endpoint, never assume the preferred port was bound. Clear stale inherited PMX_CANVAS_URL values in the host configuration as well as the shell.
• CLI query/mutation commands are HTTP clients and do not start a server. serve or --mcp starts or attaches to one. Use the host's authenticated preview mechanism for remote viewing; a loopback URL is only reachable inside the machine/orb running the service.

Choose the Smallest Useful Node Type

Use the lightest tier that communicates the result. Do not build a web artifact when markdown, json-render, a graph, or an HTML primitive is sufficient.

Current MCP Composites

The live MCP surface is 22 tools: the 16 composites below plus 6 standalones. This table is kept in sync with the server's composite registry; tools/list on a fresh pmx-canvas --mcp is always authoritative, independent of the installed release version.

Important routing:

• Basic nodes: canvas_node { action: "add", type, ... }
• HTML: canvas_node { action: "add", type: "html", html }
• HTML primitive: canvas_node { action: "add", type: "html", primitive, data }
• Graph: canvas_render { action: "add-graph", ... }
• JSON render: canvas_render { action: "add-json-render", ... }
• MCP app: canvas_app { action: "open-mcp-app", ... }
• Excalidraw: canvas_app { action: "diagram", ... }
• Web artifact: canvas_app { action: "build-artifact", ... }

As of v0.3.0, the 57 legacy single-purpose tools from the v0.2 compatibility window are removed. The composites above plus the retained standalones are now the whole MCP surface: canvas_batch, canvas_pin_nodes, canvas_screenshot, canvas_ax_interaction, canvas_ingest_activity, and canvas_invoke_command. Snapshots are the canvas_snapshot composite (actions save | list | restore | delete | gc | diff); the 6 legacy snapshot standalones were removed in v0.4.0 after their deprecated 0.3.x window.

Spatial Rules

• Treat proximity as relatedness and top-left to bottom-right as reading order.
• Search before adding to avoid duplicate nodes.
• Extend the current board in place; do not evict prior nodes to add new material.
• Use groups only when the frame communicates meaningful containment.
• Keep related nodes 40–80 px apart and separate unrelated clusters by roughly 150–250 px.
• Use directed edges for actual relationships, not decoration.
• Edge types: flow, depends-on, relation, references.
• After manual or batch layout changes, run canvas_query { action: "validate" }.

In-View Placement & Sizing (required for user-facing nodes)

Auto-placement (omitting x/y) is board-relative, not camera-relative: it places right of the last node or scans rows from the origin, ignoring where the human is looking. On a board with distant nodes, an auto-placed node lands off-camera.

1. Omit x/y only on an empty or locally dense board. Otherwise place near the human's attention: the pinned/focused neighborhood, or explicit coordinates beside the last user-facing output (gap ≥ 24–48 px).
2. After creating nodes the human should see, pan the camera (operating-sequence step 7b): focus for one node, fit with exactly the new nodeIds for a cluster. Use noPan (focus --no-pan) only when you must not steal the camera.
3. Never fit the whole board to "show" new work — on a board with outliers that miniaturizes everything. Always pass explicit nodeIds to fit.
4. Never leave user-facing output at far coordinates without a focus/fit.
5. Do not hand-compute a whole board's coordinates. Create the nodes (omitting x/y), then canvas_view { action: "arrange", layout: "grid" | "column" | "flow" } and finish with a fit over the new ids. Manual pixel math is what produces long-line, unbalanced boards.
6. fit sizes itself to the connected browser window (0.4.6+) — you do not need to guess width/height. Pass them only to fit for a window other than the human's.
7. Every node is a canvas card. Nodes render on the canvas where you put them — there is no docked/HUD placement and no dockPosition (removed in the rail-chrome redesign); status and context nodes are ordinary cards like the rest.

Size for content. Omitting width/height gives readable per-type defaults — prefer them: markdown 640×420, status 360×200, file 520×360, diff 640×420, mermaid 640×460, html 720×640, graph 760×520, mcp-app 960×600, web-artifact 960×720. A hosted app opened with canvas_app { action: "open-mcp-app" | "diagram" } — including the Excalidraw diagram preset — is the exception: request width: 720, height: 500 for its standard landscape preset instead of relying on the 960×600 default for a directly created mcp-app node. After the app mounts, read the node back with canvas_query { action: "layout" } and use its persisted width and height for subsequent placement or layout; the mounted app may have changed the requested dimensions. Since 0.4.6 the server clamps explicit creation sizes UP to per-type readability floors (e.g. markdown 360×180, graph/json-render/html 420×280, mcp-app 480×320) — a tiny probe size silently becomes the floor. strictSize: true is the only opt-out (a fixed scrolling frame you genuinely want small). canvas_query { action: "validate" } additionally reports any node below its floor as an advisory sizeWarnings entry — treat a non-empty list as layout work left to do.

Token hygiene. For routine state checks use canvas_ax_state { action: "get" } WITHOUT includeContext — the full AX context payload is ~10× larger; request it only when you are actually consuming context.

Context Pins

Context pins are the primary human-to-agent bridge:

1. Human pins nodes in the browser using Pin as context.
2. Agent reads canvas://pinned-context.
3. The resource includes pinned nodes and nearby unpinned neighbors.

Do not confuse context pinning with Lock position, which only excludes a node from auto-arrange. Every node type, including status, can be removed through canvas_node { action: "remove" }, the title-bar × control, or the Close context-menu action.

Sessions & the human

The board has three modes, all gated on one fact — whether a session is attached:

• Quiet board (nobody attached, nobody writing): a plain canvas. Your writes still show as ghosts, nothing else changes.
• External steering (you write with no session attached): the top bar shows a passive writers indicator with an activity feed listing each write under your label, plus a connected-writers sheet. Identify yourself (PMX_CANVAS_AGENT_SOURCE, or x-pmx-source on HTTP) so the feed names you, not the transport. Pending explicit intents carry an inline Veto there.
• Focus session (attached): your cursor sits on the node you last touched with a phase chip (idle / thinking / tooling / waiting-approval); every MCP/HTTP write is attributed to your session automatically (pass agentId only to keep a sub-agent separate); the human gets the session panel (work items, approval gates, timeline of your tool runs, board writes, evidence and steering), a command bar that posts steering you read on your next turn (with a target picker when several agents are connected), a timeline filterable by kind (Updates / Steer / Events / Evidence), and a context meter. That meter is the pinned-context payload estimated against a configured budget ("Pins") unless your host reports your real token usage on the presence (set-presence { contextUsage: { used, total } }) — then it shows your actual window ("Context"). The Copilot extension reports it automatically; other hosts report it themselves. Report phase: "thinking" before a long reasoning stretch if your host gives you a hook; tooling is derived from your writes. Hosts with adapters (the Copilot extension) attach for you; the human can also start a session from the board's Start agent session button — your writes (transport or PMX_CANVAS_AGENT_SOURCE label alike) are attributed to it and it takes your name.
• No adapter? Use the pump. Any CLI agent becomes steer-reactive with one command in a terminal: pmx-canvas pump --consumer <your-key> --exec '<command>'. It long-polls your delivery queue, runs the command once per steer (message on stdin plus the PMX_STEER_MESSAGE/ID/SOURCE/TARGET/CREATED_AT env envelope), and marks per-consumer only after the command exits 0 — a failed hand-off stays pending and the pump exits non-zero. {message} in the template expands to a quoted env reference (never spliced; refused on Windows — read stdin there). --parent <key> rolls you up under an orchestrator's chip; --once for scripts. See pmx-canvas pump --help.

What the session asks of you:

• Detach explicitly (attached: false or a session-end activity). Attaching over a non-empty board saved a Before session · … snapshot; detaching emits the receipt (items done / vetoed, a diff against that snapshot, one-click restore) — an idle timeout delays it.
• 403 = outside the scope fence. The human may fence you to a region (policy.scope): writes outside it are refused with a reason naming the node or position (a fenced group frame grants its members too). Read policy.scope in canvas://ax-context, ask the human to widen it, never retry blindly. The fence is the human's: set-policy does not take scope.
• 409 = a human is holding that node (dragging or editing it right now). Requeue the change and retry in a moment. If you had signalled an intent on it, it was vetoed — a yield timeline event says who took over.
• A gate you do not answer auto-holds (held, default TTL 5 min). held is a non-approval: do not proceed; the human can reopen it from the panel. A rejection reaches you as steering.
• Your latest edit can be undone from the panel. One shared undo stack: when the human undoes it you get steering ("Undid your edit: …"); treat the board as the truth, not your last write.
• Steering arrives as steering-message rows in canvas://ax-timeline (or claim them via canvas_ax_delivery). The command bar, gate rejections, undo and take-overs all speak through it. Steering may be ADDRESSED: with several agents connected the human picks a recipient in the composer, and a claim only returns broadcasts plus messages targeted at your consumer label — so always claim as yourself (your PMX_CANVAS_AGENT_SOURCE label), or addressed steering never reaches you.
• Steering is also the agent-to-agent mailbox. To coordinate with another agent on the board, send canvas_ax_timeline { action: "send-steering", message, target: "<their consumer label>" } — they claim it once on their next turn, you never receive your own, and the human sees the exchange in the panel as "you → them · message". Read canvas_ax_state { action: "presence" } for who is connected. Park shared state in work items and evidence; split territory with group frames rather than editing the same node.

Browser Workflows

Use the visible workbench when the human is actively curating layout:

• Drag nodes to move them.
• Drag empty space to lasso-select (Select tool, the default); hold Space or pick the rail's Pan tool to pan instead — the Pan tool pans even when the drag starts on a node.
• The selection bar (floating bottom-center) offers count, align left/top, distribute, auto-arrange, Group (G), Connect, Pin as context, delete and clear; selected nodes show an accent outline with corner handles.
• Double-left-click a node for context pinning, position locking, focus, collapse, connect, refresh, open, close, and type-specific actions.
• Double-left-click empty canvas for the node-creation menu; hold and drag the right mouse button to pan.
• Drop files or URLs to create matching nodes; an empty board shows starter actions (new note, pick files, paste a link, start an agent session).
• Use a markdown node's Edit button to edit its content.
• Groups are frames with the name pill and an action cluster (auto-arrange children, collapse, ⋯ rename / ungroup / pin all) on the top edge. Membership changes only on release while the "release to add to " pill shows (Esc keeps it out); dragging a child fully out offers "release to remove". A collapsed group is a chip that hides its children (edges to them draw to the chip). G groups the selection, Shift+G ungroups. Ungroup dissolves the frame (children stay, nested children move up one level, one undo step) — and your canvas_group { action: "ungroup" } is the same operation with the same result.
• Edges: drag from a node port, or pick the rail's Connect tool (C) and drag from anywhere on a node; the target lights up, Esc cancels, L asks for a label on release.
• The History drawer (rail camera button, or the receipt's Full log) lists snapshots and agent sessions in one timeline; save a snapshot there before experiments and restore only after confirmation. Ctrl/Cmd+Z and Shift+Z work the shared undo stack.
• The minimap (bottom-right of the canvas) is a true-scale map; hover magnifies it, click jumps, drag pans. A banner under the top bar reports a dropped stream (reconnecting) or a post-reconnect resync; edits still save over HTTP meanwhile.
• Other open tabs appear as green cursors with a name tag (/workbench?name=mia sets yours); a node you drag is locked for agents until you release it.
• The chrome is a persistent 52px left tool rail plus a slim 44px top bar (0.4.8+). The rail carries the tools (Select V, Pan Space, Connect C), node creation (markdown M, image I, file Shift+F, webpage W, HTML surface H, group G, annotate A — a popover with draw / text / eraser), and utilities: search & commands (Cmd+K — actions with shortcuts, then jump-to-node), arrange, trace, minimap, history, the theme picker (nine themes: dark, light, high-contrast, midnight, sepia, arctic, ember, forest, volt), and shortcuts (?). The top bar holds the connection dot, workspace title, the session chip / gate badge / context meter while a session is attached (or the external-writers indicator and Start agent session otherwise), and the zoom cluster (zoom out, % label = reset, zoom in, fit F). Hovering a rail button shows a tooltip with its shortcut. On viewports ≤1180px the top bar drops its meta text and the session panel becomes a drawer; every control stays in the rail at any width — there is no separate mobile menu.
• Keyboard: nodes are focusable — arrow keys move to the nearest node, Enter opens it in focus mode (a scrim + inset view with Open in tab), Esc closes the top-most overlay.
• Embedding hosts can open /workbench?theme=<name> (or ?theme=auto to follow the host's light/dark appearance) for a session-local default theme that never changes the server-global theme other clients see; an explicit pick from the theme menu ends the override. The bundled Copilot extension opens its panel with ?theme=light.

After changing files under src/client/, rebuild with bun run build before manual browser verification.

AX Interactions

Node interactions request PMX AX primitives; they never execute arbitrary shell, tools, MCP calls, or host actions.

• DEFAULT_NODE_AX_CAPABILITIES is the per-node-type ceiling.
• data.axCapabilities may enable or narrow capabilities but cannot escalate beyond the ceiling.
• Sandboxed surfaces are scoped to their own source node.
• HTML nodes must explicitly opt in.
• Use window.PMX_AX.emit(type, payload) and await its result.
• Listen for pmx-ax-update when an HTML control surface reflects live AX state.
• Steering is queued; claim with canvas_ax_delivery, act, then mark delivered.

Read AX HTML control surfaces before building an interactive AX-enabled HTML node.

Ready-made control surfaces (0.4.7+) — prefer these over hand-rolling

Two HTML primitives ship AX-enabled, so you do not need to author a control surface or pass axCapabilities yourself. They are the ONLY primitives created with AX on; every other kind is still a static document.

• ax-board — task list: create tasks, watch status change live, steer, bounded loop.
• ax-flow — the same controls drawn as a task flow with a loop-back rail, plus Materialize to board.

ax.flow.materialize is the one interaction that creates canvas nodes, and it is deliberately narrow: you supply TEXT ONLY (<= 12 steps, title <= 120 chars) and the server owns the result — one markdown node per step, flow edges between them, a dashed references loop edge, and one work item per step linked to its node. Re-materializing REPLACES the previous flow (the source node keeps a manifest of what it created), so it is safe to call repeatedly.

The flow also works without the panel. Materialized step nodes carry data.axStep, and the canvas renders native Start / Done / Blocked controls on them, with Run loop / Stop and a steer box on the anchor (first) step. Those controls also follow the node into focus mode. The native loop runs server-side: it advances when a step's work item completes, survives a browser reload, and keeps going while the tab is closed — unlike the panel's loop, which dies with the iframe. Bounds are the same either way: advances only while running, hard cap 20 runs, Stop persists immediately, and blocked/cancelled halts it.

When you want the agent to work through a checklist the human can watch and steer, reach for ax-flow + materialize rather than a markdown to-do list — the status chips and the loop come free.

Working a flow: keep the board honest as you go

A materialized flow is only useful if it tracks reality. While you work one:

1. Move the step to in-progress BEFORE you start it, and to done when it is actually done — not in a batch at the end. The human is watching the chips to know where you are; a board that only updates on completion tells them nothing while it matters.
2. Pin the in-progress step as context (canvas_pin_nodes with just that node, or POST /api/canvas/context-pins mode set). This makes the active step the one thing in canvas://pinned-context, so the board's "what am I working on" and your own context are the same answer instead of drifting apart. Re-pin as you advance.
3. Mark blocked rather than going quiet when you are stuck — a stalled in-progress chip is indistinguishable from a crashed agent.
4. Read steering back with canvas_ax_delivery as you go; a human watching a live flow steers mid-run, and a steer you never claim is a correction you ignored.

Resources

Read the smallest resource that answers the question:

• canvas://pinned-context — curated context plus neighborhoods
• canvas://summary — compact board overview
• canvas://layout — complete state
• canvas://spatial-context — clusters and reading order
• canvas://history — mutation history
• canvas://code-graph — detected file dependencies
• canvas://ax-context — compact AX context
• canvas://ax-work — work items and gates
• canvas://ax-timeline — events, evidence, steering
• canvas://ax-pending-steering — adapterless delivery queue
• canvas://skills and canvas://skills/<name> — bundled skills

Prefer canvas_query { action: "search" } over parsing the full layout.

Known Limitations

• Hosted MCP-app/ext-app nodes such as Excalidraw require the in-canvas host bridge and are not standalone Open as site targets. URL-backed viewers and bundled web artifacts remain openable.
• A standalone html surface (/api/canvas/surface/:id opened as a site) is a VISUAL view: it renders the same content and theme, but window.PMX_AX is not injected without the canvas iframe's per-mount nonce, so AX buttons only work inside the in-canvas node (0.4.4 Codex note). Do not tell a user a standalone tab's controls will steer the agent.
• A hosted ext-app (Excalidraw) node in a WebKit host panel (e.g. the GitHub Copilot app's embedded WKWebView) historically could render as a black tile — a host compositor paint race on the nested iframe, not a broken node (the session is healthy, sessionStatus is ready, and it renders fine in Chrome/Codex). Since 0.4.6 the canvas runs a paint oracle + recovery ladder under WebKit: present-at-load ext-apps mount strictly one at a time (the cold burst was the trigger), each frame answers a double-rAF paint probe after settle, and on silence the ladder escalates — soft-expand cycle (the automatic analogue of the proven enlarge+close) → serialized remount → an explicit "App surface failed to paint / Retry" affordance. The connecting overlay stays up until paint is confirmed, so a black layer is never presented as ready. The recovery trail (GET /api/canvas/debug/ext-app-recovery / window.__PMX_EXTAPP_LOG) now records mount-slot, paint-ok, paint-fail, soft-expand-cycle, recovery-exhausted, and assume-visible-rearm — when diagnosing, trust paint-ok/paint-fail, and never assert health from settled alone. Exception: in a host that reports the document hidden (the GitHub Copilot panel does so continuously), paint-ok is recorded as paint-ok (unverified: host hidden) — the app answered the paint probe, but nothing is proven composited, so do not treat it as a verified paint. If a tile still shows the Retry affordance, click Retry (fresh recovery budget) or expand-then-close; attach the recovery trail when reporting.
• Ext-app frame documents live in server memory. Through 0.4.0, killing/restarting the daemon while a panel stays open leaves ext-app tiles on dead frame URLs (Frame document not found, 0.4.0 report Finding S) until a full workbench reload. Since 0.4.1 the browser revalidates its frame documents on every reconnect and re-mints them against the new process automatically — if a post-restart tile still looks blank on an older install, reload the workbench page.
• A hosted ext-app (Excalidraw) resized NARROW/TALL (e.g. 360x529) can show its diagram in the upper region with the app's own dark fill below it — in every engine (0.3.4 report Finding Q). This letterboxing lives inside the hosted app bundle's root container, not in PMX (a body-level background override ships but cannot reach the app's inner root). Keep ext-app tiles landscape — at or above the 720x500 they open at, or the same ratio — or expand the node: the fullscreen overlay renders full-bleed. The durable fix is upstream in the excalidraw-mcp app.
• Behind proxies that buffer streaming responses (e.g. portal hosts), the workbench auto-falls back from SSE to a polling transport within ~3s, so the board still boots and stays live. Force a mode with /workbench?transport=poll (or transport=sse) when diagnosing.
• Nested-iframe embeds (e.g. the Amp orb portal renders the canvas page inside an ampcode.com iframe) can block child iframes from loading ANY src URL, breaking every iframe-backed node with a gray placeholder. The canvas probes this at boot and auto-falls back to fetching same-origin surfaces and rendering them inline via srcdoc (HTML, graph, json-render, frame documents). In Amp orbs specifically, the server sees AMP_ORB in its environment and adapts end to end: it binds the portal-assigned $PORT automatically (no port flag in the service command), stamps the page so the embedded client skips the (there-unreliable) probe, goes straight to srcdoc, AND defaults to the polling transport (the portal proxy buffers SSE — waiting out the watchdog could trip the boot modal); HTML surfaces also inline their theme stylesheet so they render styled. External app URLs cannot be inlined (cross-origin) and may stay blocked in such hosts. Force a mode with /workbench?iframe-mode=srcdoc (or iframe-mode=src) when diagnosing.
• Graph and json-render standalone surfaces use display=site and fill the browser viewport, and reflow on a live window resize. Some single-tab host browsers historically didn't deliver live-resize events; current Codex builds do — the 0.4.6 pass watched a standalone graph reflow live (SVG 1550×783 → 850×483 on a 1600×900 → 900×600 window resize) with no reload. Do not present a system browser as a workaround for stale resizing; it is only a preference for viewing a full page beside the canvas.
• Some hosts cannot automate inside sandboxed workbench iframes. Verify those interactions in a system browser or through server-side AX state.
• pmx-canvas screenshot requires an active WebView. Start it with canvas_webview { action: "start" }.
• The default server port is 4313, but it may fall back or be explicitly changed.

Persistence

State lives under .pmx-canvas/, primarily in canvas.db. It includes viewport, nodes, edges, annotations, pins, snapshots, AX canvas state, and large-node blobs.

• Stop the server or close/flush the SDK before committing canvas.db.
• History is session-scoped and is not persisted.
• Timeline AX data persists independently from canvas snapshots.
• canvas_view { action: "clear" } clears canvas-bound state but not host/session diagnostics.

Detailed References

Load only the reference relevant to the task:

• Full MCP, HTTP, CLI, layout, and workflow reference
• Installing PMX Canvas
• HTML primitives
• Excalidraw diagram authoring
• AX HTML control surfaces
• GitHub Copilot adapter
• Codex app adapter

The authoritative current MCP inventory and legacy replacement table is docs/mcp.md.