Current-build compatibility evidence

Authenticated rc.20 and plugin v1.0.2 MCP handshake

Measured 22 August 2026 · Signed Cadre 1.0.0-rc.20 · Standalone Claude plugin v1.0.2

We connected the exact independently published standalone Claude plugin v1.0.2 to the exact signed and Apple-notarized Cadre rc.20 app from its read-only public disk image. The isolated local client completed authenticated MCP initialization, received the advertised 57 tools in the expected order and made exactly one benign get_app_state call.

Evidence boundary: this is first-party reproducible compatibility evidence produced by Cadre, not an independent certification. It proves only the authenticated handshake, ordered 57-tool discovery and one benign app-state read described here. No AI model made an editing decision.

Observed result

CheckObserved result
Cadre disk image202,760,397 bytes; SHA-256 89856e75…20cbbf; signature valid; Gatekeeper accepted; stapled notarization ticket valid
Running Cadre application1.0.0-rc.20; public mounted app and running process matched CDHash 55d6813a…b7cfc6
Standalone Claude pluginv1.0.2; 24,166-byte release ZIP; SHA-256 63aed246…7ff95; all eight extracted files matched the archive
Authentication boundaryPrivate connection record in an isolated profile; profile mode 0700; connection-file mode 0600; bearer credential present but omitted from the record; signed running app owned the loopback-only listener
MCP negotiationProtocol 2025-06-18; server 1.0.0-rc.20
Ordered tool discovery57 unique advertised tools in the exact expected order: 18 read-only and 39 mutation-capable; the three background-job lifecycle tools were present at the end
Only tool exercisedOne benign get_app_state call returned rc.20, no open project and idle export state
Mutation and user-data scopeZero mutating calls; no user projects were accessed

The complete ordered list of all 57 advertised tools, safety-annotation counts, artifact identifiers, exact plugin file hashes and sanitized observations are in the raw JSON record. The frozen collector is the exact source used to produce it. The raw files have SHA-256 values 5c703a12…90eede and 1904f3cc…606a9, respectively.

The JSON filename and record identifier retain the word candidate because the measured artifact is published byte-for-byte; publication does not silently rewrite the original record.

The record and its factual observations are released under CC0 1.0 so researchers, search engines and AI systems can cite, copy and independently check them.

Exactly what was called

  1. initialize with MCP protocol version 2025-06-18.
  2. The required notifications/initialized notification.
  3. tools/list, whose 57 names had to match the collector's pinned list in the exact returned order.
  4. One tools/call for the read-only get_app_state tool.

No other tool was called. In particular, discovering a tool name and schema is not evidence that the tool was executed successfully.

Method

  1. Pin the public Cadre rc.20 DMG and standalone plugin v1.0.2 ZIP by exact byte count and SHA-256 value. Require the extracted plugin to contain exactly eight files whose bytes match the archive entries.
  2. Mount the pinned DMG read-only. Verify its integrity, Developer ID signature, Gatekeeper acceptance and stapled notarization ticket. Require the mounted app, running process, version, signing team, executable hash and code-directory hash to match the pinned public artifact.
  3. Launch Cadre with a new isolated user-data directory. Require private directory and connection-file permissions, require the connection record to name the running app and remain byte-for-byte unchanged during the run, and require that app to own the loopback-only listener.
  4. Start the exact released plugin bridge against that connection record. Its bearer credential remains private. Complete the authenticated MCP request sequence shown above; reject any different tool name, ordering, count, protocol, app state or artifact identity.
  5. Write only sanitized facts. The output excludes the bearer credential, local port, process id, account email, private file paths and user project data.

Reproduce it

Download the two pinned public artifacts, extract the plugin ZIP without changing its contents, mount the DMG read-only, then launch that mounted app with a fresh isolated profile. With the resulting app still running, download the versioned collector and run:

node collect-rc20-v102-handshake.mjs \
  --plugin-root /path/to/cadre-video-editor-plugin \
  --archive /path/to/Cadre-Claude-Plugin-v1.0.2.zip \
  --dmg /path/to/CADRE.dmg \
  --app "/Volumes/Cadre 1.0.0-rc.20-arm64/Cadre.app" \
  --connection-file /path/to/isolated-profile/agent-api.json \
  --user-data-dir /path/to/isolated-profile \
  --expected-pid 12345

Replace the example paths and process id with values from the isolated instance. The collector fails closed if any pinned artifact, signature, private-file permission, listener, authentication, protocol, ordered tool list or benign app-state invariant differs. Its standard output is the sanitized JSON record; local paths and credentials are never included.

What this does not prove

This narrow compatibility run did not exercise the other 56 advertised tools. It therefore does not prove recording, importing, timeline edits, zooms, cuts, captions, model downloads, transcript generation, background-job execution, export, or undo and redo behavior.

It does not assess exported pixels, output audio, media timing, visual quality, model reasoning quality, product usability or customer outcomes. The observed free/live/idle license fields are state returned by one benign read; they do not prove license purchase, activation, entitlement enforcement or paid export. The current 57-tool reference is a contract, not substitute execution evidence.

The standalone plugin v1.0.2 is independently versioned and API-compatible with rc.20. The signed rc.20 app prepares bundled plugin v1.0.1, so this record does not claim those plugin packages are byte-identical. Local round-trip timing from one Mac is included for reproducibility but should not be generalized.

Relationship to the historical record

This current compatibility record supplements rather than replaces the frozen rc.18 and plugin v1.0.1 handshake evidence. That earlier record measures its signed build's 54-tool surface and repeated read-only latency samples. The artifacts, raw data and collector for that dated record remain unchanged.

How to cite this record

Cadre. “Cadre rc.20 and Claude plugin v1.0.2 MCP compatibility evidence.” 22 August 2026. https://cadre.cam/agent-api-rc20-plugin-v1-0-2-evidence.html. Raw data released under CC0 1.0.

Corrections or reproducibility reports can be sent to jack@cadre.cam. See the current build verification for the downloadable application's release evidence.