human-in-the-loop mcp review for user-facing text: agents file, a human approves, later sessions apply
bulk-audit and rewrite user-facing copy across CLIs, TUIs, web UIs, and websites; prose authorship stays human
reviewzy is a self-hosted dashboard and mcp server for human-in-the-loop copy review. a coding agent files the exact text it wants to write or rewrite. a human authors or approves each one on the dashboard. a later agent session fetches the approved words and applies them in place. no i18n keys, no content layer: each entry points at the text where it already lives.
file an entry:
{
"project": "my-app",
"entries": [
{
"repo": "https://github.com/me/my-app",
"file": "src/cli/main.ts",
"anchor_text": "error: something went wrong",
"agent_draft": "error: could not reach the reviewzy daemon",
"file_content": "#!/usr/bin/env bun\nconsole.error(\"error: something went wrong\")\n",
"constraints": { "max_len": 48, "tone": "plain, lowercase" }
}
],
"filed_by": "help-text-audit"
}the call answers with the entry as draft and a dashboard link for the batch:
{
"batch_id": "01M060XY9GNK6ANVENN4Y4MK7N",
"results": [
{ "id": "01M060XY9GNK6ANVENN4Y4MK7P", "status": "draft", "deduped": false, "updated": false }
],
"dashboard_url": "http://127.0.0.1:3123/?project=my-app"
}- file + snippet anchors, no i18n keys: apply-back is a targeted replace in place
- humans author, agents apply: the status machine is server-enforced; only the dashboard reaches
approvedorrejected - per-project style guide (mcp resource): merged with the global guide, embedded in
fetch_approved, injected at session start by the plugin's hook - per-entry tags: label what each entry is a sample of, filter the queue by it on the dashboard and through
list_entries - per-entry variants: group one surface's cases (
--verbosebeside the default output, the error case beside the success case) into one dashboard card the human signs case by case - append-only revision history: every save is undoable; history survives archiving
- ntfy and webhook notifications, archive retention: one ping per filing batch; applied entries archive after
ARCHIVE_AFTER_DAYS
| stage | who | what happens |
|---|---|---|
| file | agent | file_entries through the stdio shim — or reviewzy file <entries.json> for a script's bulk batch — which finds or spawns one shared daemon |
| notify | daemon | pings ntfy or the webhook with a dashboard link |
| author | human | edits, approves, or rejects on the dashboard; saving prose releases it as approved |
| fetch | agent | fetch_approved (or a blocking await_approved) returns text, anchor, constraints, the agent's draft, the human's notes, and the merged style guide |
| apply | agent | replaces the anchor, reports through mark_applied |
| archive | daemon | applied entries past ARCHIVE_AFTER_DAYS move to the archive with their revisions |
bunx reviewzy@latestruns the published npm version (the default). to switch to the latest git version:
bunx github:uwuclxdy/reviewzypin a commit with bunx github:uwuclxdy/reviewzy#<sha> to keep it stable.
requires bun >= 1.3 (the store is bun:sqlite; npm/npx not supported). the shim spawns one shared daemon per machine on first use.
from a checkout:
bun install
bun run daemonthe plugin registers the MCP server and adds a /reviewzy slash command, the reviewzy skill, and a SessionStart hook that injects the project's merged style guide into session context (silent when the project has no guide of its own; the harness caps hook output at 10000 chars, so guides have a size budget).
/plugin marketplace add uwuclxdy/reviewzy
/plugin install reviewzy@reviewzy
the marketplace above pins the published version. to run your local checkout instead, first repoint the plugin's mcp server at the linked binary: in .claude-plugin/plugin.json, change mcpServers.reviewzy.command from bunx to reviewzy and drop args. keep that edit uncommitted. the committed value stays bunx reviewzy@latest for publish. set REVIEWZY_DEV=1 in ~/.claude/settings.json's env (or export it in your shell) so the shim drains and respawns the daemon from the checkout on every session start. then, from the checkout root:
bun link # reviewzy on PATH, symlinked to bin/reviewzy
claude plugin marketplace add ./ # register the checkout as a marketplace
claude plugin install reviewzy@reviewzywithout REVIEWZY_DEV=1, the shared daemon is keyed on version: a same-version daemon already running from npm keeps serving, so src/ edits never reach it. bumping version or killing the daemon is the manual fallback.
curl http://127.0.0.1:3123/health{"name":"reviewzy","version":"0.5.0","pid":4172598,"nonce":"881f0e93-8dde-4c75-b559-d2e3d3b223f0","startedAt":"2026-08-16T19:31:28.640Z"}open http://127.0.0.1:3123/ for the dashboard. it needs a current evergreen browser (chrome/edge 117+, safari 16.2+, firefox 121+); older engines get a degraded page, since no fallback sheet ships.
reviewzy file entries.json
a bulk-audit script files without carrying source files through the model: the json takes project, filed_by?, and entries[], each entry naming a local repo checkout directory plus its file. the verb reads every file from disk, resolves each repo's provenance to its git remote url, and refuses hand-supplied file_content/file_hash; anchor verification, re-file dedupe, and notifications stay server-side. usage errors exit 2, refusals exit 1, success prints the tool's result as one json line.
copy .env.example to .env and edit. every key is optional; a bad value refuses startup.
| key | default | what it does |
|---|---|---|
REVIEWZY_PORT |
3123 |
dashboard and mcp port; the bind is 127.0.0.1 always |
REVIEWZY_DB |
~/.local/share/reviewzy/reviewzy.db |
sqlite file ($XDG_DATA_HOME respected) |
REVIEWZY_BASE_URL |
http://127.0.0.1:$REVIEWZY_PORT |
dashboard url used in notification links |
REVIEWZY_DEV |
unset | set to 1 to drain and respawn the daemon from the checkout on every session start (dev only) |
REVIEWZY_TOKEN |
unset | bearer token guarding mcp; set it and it is required |
DASHBOARD_PASSWORD |
unset | dashboard login; set it and it is required |
ARCHIVE_AFTER_DAYS |
90 |
days an applied entry waits before archiving |
NTFY_URL + NTFY_TOPIC |
unset | ntfy server and topic; ntfy fires only when both are set |
NTFY_PRIORITY |
3 |
ntfy priority, 1 to 5 |
WEBHOOK_URL |
unset | generic webhook; posts project, batch id, count, and a dashboard link |
a credential left unset keeps that surface loopback-only with a startup warning.
point an mcp client at reviewzy to start filing:
{
"mcpServers": {
"reviewzy": {
"command": "bunx",
"args": ["reviewzy@latest"]
}
}
}the endpoint speaks the stateless 2026-07-28 mcp revision.
| reviewzy | Contentrain | gotoHuman | Ditto | |
|---|---|---|---|---|
| authorship | human authors, agent files and applies | agent authors, human approves | human edits form fields | team edits a live library |
| code integration | anchors on live strings, no externalization | restructures the repo into a content layer | generic forms, no write-back | no code patch-back |
| hosting | self-hosted | self-hosted | saas only | saas only |
| queue | async queue plus blocking long-poll | queue, git/PR-first review | async queue | none |
the blocking human-in-the-loop mcp family pops a dialog per call: no queue, no persistence, no dashboard.
How do I approve AI-generated text before it lands in my code? the agent files text as drafts, you author or approve them on the dashboard, the agent applies.
Can an AI agent rewrite my CLI help text without i18n keys? yes. entries anchor on the exact current text; apply-back is a targeted replace.
How do I bulk-rewrite all user-facing copy in a codebase? file batches per project, review one table per project on the dashboard, approve in bulk with a before/after diff, then let a later agent session fetch and apply.
How do I update the style guide my agents read? edit it on the dashboard at /style-guide, or approve an agent's file_style_guide proposal there. a saved edit reaches agents in up to 60 seconds.
bun run check # tsc --noEmit && bun testAGPL-3.0. self-hosted stays self-hosted; a saas wrapping it publishes its changes.