Webflow Marketplace / 06 of 08
Bring review context to the reviewer’s tools
I built review tooling that gives an AI assistant structured access to Marketplace queues, asset versions, prior feedback, and submission evidence, with explicit limits on changes.
Selected professional work by Micah Johnson for Webflow, presented as individual portfolio work. Webflow did not commission this work through CREATE SOMETHING. Individual contributions are described within a wider team and platform context.
Contribution focus: Template and app review integrations, evidence retrieval, and write boundaries
MCP is the Model Context Protocol, which connects an AI assistant to structured tools and information.
Offline evidence replay with synthetic clients. Actual handler results; no live reviewer or Airtable operation.
Silent recording · Select captions for visual guidance. Open the MP4 ↗
Recorded use / 30-day snapshot
The work behind
the numbers.
4 October–2 November 2026 · UTC
Checked 3 November 2026 · Cloudflare D1A dated view of activity in the tools. Select a measure to inspect what was recorded and what it can establish.
Review toolsRecorded tool calls
All definitions are available in the counting method below.
What this records
Template Review MCP
Calls to tools that retrieve template review context, run checks, and support review actions.
87 tool failures were recorded within these calls. A tool call is one step in a review, not a completed review.
Cloudflare D1 · tool-invocation records
This count does not establish review quality or final approval.
Tests, operator activity, repeat visits, and retries are not separated. These counts do not establish unique users, completed reviews, or hours saved.
Counting method and full definitions
The window begins on 4 October 2026 at 00:00 UTC and ends before 3 November 2026 at 00:00 UTC. This is a fixed snapshot, not a live counter.
- Template Review MCP · 9,188 recorded tool calls
- Calls to tools that retrieve template review context, run checks, and support review actions. 87 tool failures were recorded within these calls. A tool call is one step in a review, not a completed review. This count does not establish review quality or final approval.
- App Review MCP · 959 recorded tool calls
- Calls to tools that bring app versions, review context, and submission evidence into the reviewer’s workflow. No tool failures were recorded in this window. Recording is best effort, so that is not a guarantee of perfect reliability. Tool results do not establish review quality or final approval.
Counts come from stored tool-invocation records and legacy Dashboard analytics in Cloudflare D1. Recording is best effort. These records are separate from the historical footage and offline replays. They do not establish revenue or a measured improvement over an earlier process.
The problem
Template/App Review MCP
A reviewer needs to understand the current version and prior decisions before deciding what to do next. Reconstructing that context across operational systems adds work, and an assistant with broad write access can introduce a second problem: acting on incomplete or stale information.
The contribution
The work behind the interface.
The Template and App Review MCPs expose domain-specific review tools. They bring record context into the reviewer’s agent workflow while retaining important differences between template and app operations.
Template self-assignment uses runtime reviewer identity. Supplemental agent notes have their own feedback field. Confirmed app status changes require an expected current status, so stale state or missing confirmation can fail explicitly. App reviewer assignment remains read-only through this interface. Preflight evidence is available as review context rather than an approval shortcut.
The key decision
What needed to stay clear.
Make the authority to act as explicit as the evidence. Tool access, reviewer ownership, confirmation, and the freshness of the underlying record all affect which action is valid.
The documented result
What the evidence shows.
The new offline evidence replay shows eight actual production-handler calls with synthetic in-memory clients: queue and context reads, confirmation and stale-state refusals, reviewer self-assignment, the review contract, and supplemental-only notes. Assertions passed with zero network attempts or external writes.
The recording is an offline replay of real handler output. It does not demonstrate live MCP transport, reviewer authentication, Airtable retrieval, or production review operations.
Recording and replay evidence captured October 2026. Historical implementation and recordings do not establish the current state of every production system.