mcp-vm-relay

The Model Context Protocol front end of the VM relay, packaged as a Claude Code plugin and as a pi package. pi-vm-relay, which gave pi a native relay tool directly, is retired; mcp-vm-relay replaces it for both Claude Code and pi, the latter loaded through pi-mcp-adapter. It is an MCP server: eighteen relay_* tools, each with its own schema, title and annotations, and the server's instructions. The user command, status, ships as one small command file per host (see "Commands"). There is no skill and no hook. Together they offer recorded, snapshot-evidenced interruptive computer-use and browser-use in a fresh, dedicated VM. The agent judges interruption; non-disruptive work stays with the agent's local tools. Nothing here executes local computer-use, spawns subagents, targets physical machines, records video, or attaches to existing browsers.

How it is built

The relay implementation lives in this repository: the manager (src/manager.ts), the vm-service client, the registry with its OS-level lock, the guest transfer and transport, the evidence package and its state merge, the selected-environment profile, the console and saved-image contracts, the strict tool contract (src/schema.ts), the action dispatch and bounded result rendering (src/surface.ts), and the two guest programs (src/guest/receiver.ts, src/guest/mcp-host.ts). It was inherited from pi-vm-relay at commit 8990123, synchronized with its implementation through commit 0d69fc7 before that project's retirement, and is maintained here as mcp-vm-relay's own implementation from then on. src/server.ts binds the core to MCP over standard input and output.

The committed dist/ holds server.mjs (the MCP server with the core bundled in), the guest bundles the manager stages (receiver.mjs, mcp-host.mjs) and the doctor. Compiled relay-driver code, the recorded-execution and evidence substrate, is bundled with hashes and provenance in dist/build-info.json. Consumers load dist/; no build, SDK checkout or network fetch happens at install time.

The Claude Code plugin and pi both reach this same server as plain MCP: pi through pi-mcp-adapter (see "Use with pi" below), Claude Code by starting it directly. Every tool's identity, schema and contract are shared; the plugin adds nothing Claude-Code-specific beyond the marketplace packaging and the vm-relay-operator agent definition.

Install

Use with Claude Code

Try it from a checkout:

claude --plugin-dir /path/to/mcp-vm-relay

Or add this repository as a marketplace and install the plugin from it:

/plugin marketplace add WeZZard/mcp-vm-relay
/plugin install mcp-vm-relay@mcp-vm-relay

Once loaded, the model sees the eighteen tools as mcp__plugin_mcp-vm-relay_relay__relay_search, mcp__plugin_mcp-vm-relay_relay__relay_probe, and so on through mcp__plugin_mcp-vm-relay_relay__relay_status (see "The tools" below for the full table). The vm-relay-operator agent definition allows them by those names. The server can also be configured directly in a project's .mcp.json with node /path/to/mcp-vm-relay/dist/server.mjs, in which case the tools are mcp__relay__relay_search and so on.

Use with pi

pi install npm:@wezzard/mcp-vm-relay

This needs pi-mcp-adapter installed. The tools appear as relay_search, relay_probe, and so on through relay_status, with no host-specific prefix (pi-mcp.json sets toolPrefix: "none"). The package also ships a pi prompt template for the command, /mcp-vm-relay-status.

Use with npx

npx -y @wezzard/mcp-vm-relay

This runs the MCP server directly on stdio, for a client that speaks MCP without a Claude Code plugin or pi package wrapper.

Prerequisites

Node 22+, Python 3 with fcntl for the registry lock, and for VM execution an Apple-silicon Mac with Tart, a running vm-service and prepared guest images. Check readiness with node dist/doctor.mjs. It does not acquire a VM or establish guest UI readiness. Guest execution needs Node and CuaDriver with capture/input permissions; Linux requires native X11 and serve --no-overlay. The playwright and chrome-devtools targets of relay_run need a browser in the guest (installed Google Chrome, or a path given as browserExecutable); their pinned servers are downloaded once on the host and staged, so the guest needs no network or npm. Console viewing needs the vm-service guest-sharing backend and its guest preparation; see docs/console.md here.

The tools

Each tool's inputSchema is a plain object with no root anyOf/oneOf/allOf and no action field, derived directly from the strict per-action contract in src/schema.ts (nested unions inside a property, such as relay_image's target, are unaffected). A call is mapped back to that contract's shape and dispatched exactly as the retired single relay tool was, so behaviour, results, image blocks and isError semantics are unchanged from before this split. relay_run forwards the cua-driver, Playwright MCP and Chrome DevTools MCP tool calls a model already knows; relay_exec, relay_script and relay_code share the optional reason, step, snapshots and timeoutMs. Evidence is automatic for all four: see "relay_run" below.

readOnlyHint/idempotentHint are true for relay_search, relay_probe, relay_acquisition_capabilities, relay_console_resolve, relay_image and relay_status; destructiveHint is true only for relay_finish and relay_release, which destroy the VM; openWorldHint is true only for the four run tools, whose guest code may reach the network.

A run's result carries the execution identity and outcome, the derived step record, and, when the snapshot plan captured the after phase and delivery succeeded, the saved after-image as an MCP image content block. A command's result also carries the guest's bounded standard output and error; a relay_run result carries the target tool's own text and image blocks. The full receipt stays in the evidence package.

Commands

The command asks the assistant to call one relay tool and report the answer. It changes nothing itself.

It is not an MCP prompt. pi's MCP adapter can only name a prompt /mcp__<package>__<server>__<prompt>, so each host gets its own command file instead: pi prompt templates in pi-prompts/ (listed in package.json under pi.prompts) and Claude Code plugin commands in commands/. Both are generated by npm run build from scripts/host-commands.mjs; do not edit them by hand. CI and the release workflow check the packed tarball with scripts/verify-package.mjs: the generated files must match their source, pi's glob must select exactly the one template, and the packed server must start without the prompts capability. The release publishes and attaches the same tarball it checked.

Ownership, failure and recovery

• One server session owns at most one VM. Another task needs a separate acquisition.
• An operation failure retains the VM so the agent can inspect, repair and submit a new operation. Failed and uncertain operations are never automatically replayed.
• Use finish to deliver evidence and release, or release to abandon explicitly. A declared output that was never produced is recorded as incomplete (incompleteExtractions) rather than failing finish. A failed delivery retains the VM; a release failure retains ownership until destruction is verified.
• When the session ends (the client closes the server's standard input, a write to its standard output fails, or it is signalled), the relay renews the lease once for the rest of its TTL, then renewal pauses and the recording detaches; the VM is not destroyed and is retained until its TTL for an explicit finish or release. A finish or release already in flight completes first, for up to the shutdown grace period; after that the operation is cancelled and renewal pauses without it. The backend TTL and grace period handle abandoned leases. Status reports the last confirmed expiration time.
• While the session is live, the relay tracks the lease's own deadline (its start plus ttlHours) and renews vm-service only 15 minutes ahead at a time, never past that deadline, and stops renewing at it. A client that crashes without ending its session therefore leaves its VM leased for at most about 15 minutes, plus the backend's grace period, rather than its whole TTL.
• A restarted server with the same MCP_VM_RELAY_SESSION reconciles its durable ownership and reattaches the recording session without replaying prior work. Context compaction does not reset VM state; probe reports the owned state without relying on earlier messages.
• A run tool's timeoutMs defaults to 120,000 ms and accepts integers up to 3,600,000 ms. It bounds command execution, not the snapshot delay or the lease lifetime. A timeout reports confirmed termination or uncertainty and keeps the VM available.
• relay_exec with diagnostic: true records command diagnosis or repair without screenshots, for explicitly requested diagnosis when capture is unavailable. Before staging it runs through vm-service in the guest's default directory; after staging in the recording workspace. It cannot join a snapshot group and is not visual verification.
• After diagnosing damaged recording state, relay_stage with resetRecording: true archives the old recording and starts a new recording session in the same VM. An existing receiver lock refuses the reset until its operation is reconciled. Prior evidence paths remain in the owned status.

relay_run: MCP tool calls in the VM

relay_run takes the same tool calls a model already sends to cua-driver, Playwright MCP or Chrome DevTools MCP and runs them against that server inside the VM:

{"target":"playwright","tool":"browser_navigate","args":{"url":"https://example.com"}}
{"target":"cua","tool":"click","args":{"pid":812,"window_id":3,"x":40,"y":90}}
{"target":"chrome-devtools","tool":"take_snapshot","args":{}}

relay_tools {"target":"playwright"} returns the server's real tools/list, so the exact names and schemas are one call away. The arguments are checked against the tool's input schema (a JSON Schema validator, Ajv) before anything is sent; a mismatch is refused with the correct schema in the error.

• Evidence is automatic. Every call is admitted by the guest receiver, journaled, and bracketed by a before and an after display snapshot. The step record is derived from the call: the title is <target>.<tool> (for example playwright.browser_click), the expected result is "returns without a tool error", and cua-driver's accessibility forms are labelled accessibility. reason, expected and afterIntervalMs are optional overrides. The same rule now applies to relay_exec, relay_script and relay_code: their reason, step and snapshots are accepted but no longer required.
• Default waits before the after-snapshot: 300 ms for playwright and chrome-devtools (both servers already wait for the page to settle before they answer), 500 ms for cua and for commands.
• Sessions persist. A small guest-resident MCP host keeps one MCP SDK client per target, starting each server on first use, so a browser page or a native session survives between calls. relay_finish and relay_release stop it.
• Results pass through. The target's text blocks follow the relay's own result text, bounded like every result (50 KiB, 2000 lines); its images (up to four) are delivered inline under the usual image rules, before the relay's after-snapshot. Each call's full result, images and the servers' logs land in workspace/relay-run, the extraction relay-run, and come home with relay_finish.
• Outcomes. A call the relay proves was never sent is refused; a call that may have reached the server and has no answer (timeout, crash, a malformed answer) is uncertain, and the server is stopped so nothing it still holds can act later; the target's own isError: true is completed-with-tool-error (exit status 3). All three are error results and keep the VM. Nothing is ever replayed.

The launch commands, pinned versions and the full outcome mapping are in docs/relay-run.md.

Saved images

A run tool returns its saved after-image as a typed image block whenever the snapshot plan captures that phase: a standalone event or a text group's last event. A first or intermediate group event and a diagnostic command do not invent an image. Execution and image delivery are independent outcomes: a completed command can have a failed delivery, and a delivered image does not turn a failed command into a successful one. Either failure sets the MCP isError flag while the content, image included, is kept. The delivery identity (imageDelivery) leads the result text so it survives truncation.

If inline delivery fails, relay_image retrieves the same saved image through one closed selector; it never repeats input, creates a capture, exports the consumer directory or acquires a VM:

{"target":{"source":"display","sessionId":"<recording-session>","executionId":"<saved-execution>","phase":"after"}}
{"target":{"source":"application","name":"relay-run","path":"playwright/<call>/image-1.png"}}
{"target":{"source":"reference","imageId":"image-<64 lowercase hex digits>"}}

• Display selection needs before or after; a phase the plan did not request returns not-requested. Application selection needs an acquisition-time declaration: a directory declaration takes one relative file path, a file declaration omits path. A reference resolves to the same original bytes or an explicit failure.
• Originals are limited to 64 MiB, 40,000,000 decoded pixels and 32,768 pixels per dimension; PNG, JPEG and WebP are accepted. The preview is at most 2,000 × 2,000 pixels and 4 MiB of base64. Presentation differs from pi here: pi resizes through its own image helper, while this core has no codec dependency. A PNG original is decoded and, when it exceeds the preview bounds, resampled in-process (area averaging, node:zlib); a JPEG or WebP original has its container validated and passes through unchanged when within bounds, and is reported presentation-unavailable otherwise. No EXIF orientation is applied; dimensions are those stored. The presentation receipt records the policy, the original and preview dimensions, the MIME type, hash, size and whether bytes were transformed, beside the untouched original.
• Each delivery has a 90-second deadline and a shared budget of three byte-transfer attempts. Only transient transfer failures are retried; authorization, unsafe-path, capture, format and integrity failures are not. Recommend at most two explicit recovery calls for the same reference; if a required image still cannot be inspected, stop exploratory input and finish or release.
• Immutable catalog entries bind each reference to the owner, enclosure, backend and original identity. Closed or sealed enclosures return stale-reference; delivered originals remain readable with the host's file tools. Finalization materializes verified originals at canonical snapshot paths and preserves prior state when merging guest evidence.
• Attachment means the block was included in the result; it does not prove the model inspected it or that a human reviewed it.

Live console viewing

Acquisition never opens a viewer. relay_acquisition_capabilities reads the backend's VNC options; relay_acquire with vnc: true checks OS availability and requires a ready console and lease identity in the response. relay_console_open is permitted only following an explicit user request, requires userRequested: true, console_id, attempt_id, reason and expected, and opens on the declared service host, not on a remote client. relay_console_resolve refreshes the non-secret observation; relay_console_cancel closes managed viewing resources and never releases the VM. Console status ready is guest preflight, not launch success; transport connection, authentication, displayed pixels and human confirmation are separate observations that console actions never establish. On macOS the human must choose Standard sharing of the existing console, not a new Log In session or a High Performance display; the relay never confirms that selection, and server_enforced_view_only: false with session_binding: viewer-selection-unverified remain limitations after transport connects. Failed or uncertain launches retain ownership; resolve or cancel the attempt rather than replaying it. Console tests use mocked backends; no live viewer, installation or platform acceptance is claimed here.

Selected environments

Set VM_ENVIRONMENT_FILE to a profile that binds a loopback vm-service endpoint, an image repository, a Tart store and separate service, image and relay state directories together. The profile is validated before use, is authoritative over the individual MCP_VM_RELAY_* variables, and an invalid selection fails rather than falling back. Leases are bound to their backend and store: an owned lease restored under a different environment fails before any backend operation, and destruction is verified with the selected Tart binary and TART_HOME. docs/selected-environments.md documents the schema. The bundle the profile exports to subprocesses is the canonical one vm-service defines, so its relay entries keep the VM_RELAY_STATE_DIR and VM_RELAY_URL names and the three runtimes agree; this server reads its own settings from the profile.

Evidence and cleanup

Default output is relay-evidence/<unique-task>/ under the project, with state/ (the guest journal, action records, receipts and snapshot PNGs), host/ (reasons, submissions, transfer facts, receipts, diagnostics, image deliveries and lifecycle events), extractions/ (declared files, including the page captures), manifest.json, which checksums everything, and OPENING.txt. The package is the raw evidence only; pi-secretary's computer-use extension shows its steps and verdicts beside the agent's session. The steps and verdicts are derived from the evidence when you review it, so an earlier package is reviewed with today's rules. A program that reads relay packages gets the same derivation from @wezzard/mcp-vm-relay/review-data: await reviewData(<directory>) returns the steps (with each snapshot's package path), the verdicts and their reasons. No network, VM or external assets are needed. Diagnostic commands appear in the review data as command-only steps without screenshots. finish verifies the package before destroying the VM; a failed delivery removes only the derived files it created and retains the VM for a corrected attempt. Cleanup checks the read-only host Tart inventory (the selected one, when an environment is selected) before claiming destruction. Leases default to 4 hours, renewed 15 minutes at a time up to that TTL; the vm-service reaper is the final backstop for process death.

Relay evidence has no size limit (owner decision PS-D12): screenshots are delivered as captured, never scaled, compressed, deduplicated or budgeted, and the relay state is delivered whatever its total size or file count. Only what a client stages into the guest is bounded (512 MiB and 10,000 files per staging request), and a single image original must be at most 64 MiB to be valid. Before finish pulls the relay state it checks that the evidence volume has room for it; when it does not, finish fails before pulling, names the shortfall and keeps the VM for a retry after space is freed.

Configuration

• MCP_VM_RELAY_PROJECT: the project directory (the plugin passes Claude Code's). Evidence lands under relay-evidence/<task>/ there.
• MCP_VM_RELAY_SESSION: an explicit session identity. By default each server process takes a fresh random one; a stable identity lets a restarted server reconcile and reattach the enclosure of the same identity.
• MCP_VM_RELAY_URL: the loopback vm-service origin, default http://localhost:6240. Non-loopback servers, redirects and physical targets are refused.
• MCP_VM_RELAY_STATE_DIR: host state, otherwise $XDG_STATE_HOME/mcp-vm-relay or ~/.local/state/mcp-vm-relay, one subdirectory per session.
• MCP_VM_RELAY_REGISTRY: an explicit task registry file; otherwise an existing compatible ~/AGENTS.md VM table, or a private managed registry.md in the state directory.
• MCP_VM_RELAY_SHUTDOWN_GRACE_MS: how long a session end waits for an in-flight operation before cancelling it; default 120000.
• MCP_VM_RELAY_PYTHON: the Python 3 used for the registry lock; default python3 on PATH.
• VM_ENVIRONMENT_FILE: a selected environment profile; it overrides the URL, state directory and registry above and binds the Tart store.
• Guest files go to /var/tmp/<UTC-timestamp>-mcp-vm-relay-<task>/, with the runtime in receiver.mjs, support in support/ and work in workspace/.

Development

Build relay-driver using its own instructions, then in this checkout:

npm ci
npm run setup:dev -- /absolute/path/to/built/relay-driver
npm run check                                            # build, typecheck, tests

Three tests need a sibling checkout beside this repository and skip with a message when it is absent: the console fixture regeneration and the environment parity test need ../vm-service, and the shared catalog text corpus needs ../pilot-images. The JPEG and WebP presentation fixtures in tests/fixtures/ were encoded once from the test's spatial PNG with a codec outside this repository and are committed, because the core carries none.

End-to-end with headless Claude Code

npm run test:e2e

This runs claude --print with the plugin loaded from this checkout (--plugin-dir), the model limited to the relay's tools, and the server pointed at a loopback stand-in for vm-service (tests/e2e/fixture-service.ts), which keeps leases as records, runs the relay's own Node and transfer commands locally, and fakes the desktop capture driver. Four cases run, each a separate headless session: the status call (which also proves the server is pointed at the stand-in before anything is acquired), an invalid call refused by the contract, a full acquire, stage, exec, finish lifecycle with a verified package, and a lease abandoned when the session ends, which is retained with its renewal paused rather than released. What is checked is what happened on disk and at the service, not what the model said. The suite needs claude on PATH with an account and spends a few model turns per case, so it is not part of npm test.

The relay-driver links are development-only. The tests run the manager, the guest programs and the shipped server against fixture HTTP services, fixture MCP servers built with the MCP SDK and a real MCP client over stdio; no VM, desktop or user registry is touched. Commit the regenerated dist/ with source changes.

Releasing

Cut a release with:

node scripts/bump.mjs X.Y.Z && git push --follow-tags

scripts/bump.mjs sets X.Y.Z as the version in package.json, .claude-plugin/plugin.json and the pinned pi-mcp.json arg, refreshes package-lock.json, rebuilds dist/, commits Release vX.Y.Z and creates the annotated tag vX.Y.Z. It refuses to run against a dirty working tree or a malformed version, and it never pushes — git push --follow-tags is a separate, explicit step.

Pushing the tag triggers .github/workflows/release.yml, which rebuilds vmctl, checks the working build against a fresh one, and publishes to npm using trusted publishing (OIDC — no NPM_TOKEN secret involved), then creates the matching GitHub release.

Trusted publishing has to be configured on npmjs.com, and npm requires the package to already exist before you can do that. The first publish must therefore be done by hand (npm publish --access public from a maintainer's machine) before its trusted publisher can be configured for this workflow.

License

MIT — see LICENSE.