Portable run evidence
Reuse a bounded run snapshot, reproduce the local MCP App example, and understand verified host limits.
Two ways to carry evidence
A run's shared metadata presentation can be reused outside Studio:
- Supplied snapshot: Studio exports bounded presentation JSON. A custom host validates and renders that data without credentials or additional reads.
- MCP App: a compatible host supplies an authorized
get_runevents result and renders the linked view. Explicit refresh/next-page actions use that host's existing MCP connection for the same project/run/events context.
These are repository implementations. They do not prove that a hosted deployment has enabled agent access or that any particular AI product/version can render this view. A disabled deployment still has no working MCP connection. The shared presentation package is private workspace code, not a published embedding SDK.
Reproduce the supplied example
The public repository includes a
synthetic sample and runnable setup.
Follow its commands to build and start the local-only harness at
http://127.0.0.1:3004/evidence, paste snapshot.json, and choose
Render supplied evidence. Select a milestone to inspect its event reference.
This needs no database, ingestion key or external model call.
For your own authorized selection, use Studio → Timeline → Table → Portable evidence snapshot. Export only evidence the recipient is allowed to receive. The supplied-data host accepts at most 65,536 characters, 8 unique milestones and 4 metrics. Missing coverage or unknown fields are rejected; invalid replacement JSON clears the previous view.
A copied snapshot is historical supplied evidence. It neither authenticates its recipient nor proves current access. The example does not save, send or fetch data. It has no automatic refresh, independent query engine or analytics.
Reproduce the MCP App and fallback
The same example documents the loopback synthetic MCP fixture, exact upstream Basic Host revision, required compatibility patch and setup commands. It includes a real MCP client check that declares no UI capability and verifies complete text/structured evidence plus continuation cursors across two pages.
The MCP view renders one bounded events page, showing at most 8 event milestones.
Other get_run sections keep their full structured/text result and explain that
there is no visual projection. Protocol-compatible clients without negotiated UI support retain those
complete results, references and cursors; the UI does not replace the MCP API.
The frame receives no credentials. It delegates explicit reads through the host; each real read repeats live grant/project authorization and bounded admission. Page reductions for the response budget cannot enlarge later requests. Errors, invalid results, cancellation and closure clear evidence and read actions. Access may have expired, been revoked or changed: reconnect or select authorized evidence in the host to continue. A snapshot cannot detect unseen access changes.
Verified compatibility and limits
Consult the dated compatibility record for exact tested versions, environment, evidence scope and results. The observed MCP Apps host is the patched official Basic Host v2.0.4, using MCP peers 2.3.1 and Flowmark's 2026-07-28 protocol profile. It is a local reference-host result. Codex, Cursor, Claude and other commercial client integrations are unverified by this example. General MCP Apps support does not establish Flowmark compatibility in a particular product.
The view renders event metadata with explicit withheld fields and page omissions. It is not a redaction boundary: the authorized host can forward raw diagnostic fields to the sandbox even when they are not rendered. Use synthetic content for the reference-host smoke. Business results are application-reported, completion is not verified business success, and missing terminal telemetry does not prove that work is active or failed.
See local development, the HTTP API and security/privacy for the separate deployment, credential and evidence-disclosure boundaries.