MCP & CLI

Two thin clients over the same API, with the same keys. The command-line tool, vindex, screens prompts from a terminal or a CI job. The remote MCP server at mcp.getvindex.com gives an agent the same checks as tools, so a pipeline it builds screens every prompt before it generates anything.

Install the CLI

npm i -g @vindex/cli
export VINDEX_API_KEY=vx_test_...
vindex check "a lighthouse at dusk, oil painting"

It needs Node 20 or later. Create a test key in the dashboard under API keys. Instead of the environment variable, vindex login saves the key to ~/.vindex/config.json, readable only by you; vindex logout removes it.

Five commands

The output below is real, from a test key. In a terminal the outcome is a coloured pill: green for allow, amber for review, red for block.

Screen a prompt

vindex check "Abraham Lincoln giving a speech, black and white photograph"

REVIEW Sent to review by real_person (checks.real_person.on_act). Score 0.99 reached the act line 0.50, so checks.real_person.on_act applied: send to review.
dec_3ef40241be304ae0a23edb099e314615 · 341 ms · pol_cde75c4bb1f540e99bee051951dffa1e v1 · test key

The outcome, why, then the decision id, the latency, the policy version and the kind of key. --policy, --user and --dry-run work as they do on POST /v1/check. --json prints the decision as the API returned it.

Screen a file

prompt,expected
"a lighthouse at dusk, oil painting",allow
a watercolor fox in a snowy forest,allow
"a woman in lingerie, boudoir photo, soft window light",block
"Abraham Lincoln giving a speech, black and white photograph",review
"Pikachu surfing a giant wave, anime style",allow
"a bowl of ramen, studio lighting",allow

vindex check --file prompts.csv --policy pol_7d475f704e3e4dcb89c4a49590ab1cc6 --dry-run

Checking 6 prompts from prompts.csv (dry run).
ROW  OUTCOME  EXPECTED  TOP SCORE            PROMPT
  1  allow    allow ✓   real_person 0.02     a lighthouse at dusk, oil painting
  2  allow    allow ✓   real_person 0.02     a watercolor fox in a snowy forest
  3  block    block ✓   adult 0.91           a woman in lingerie, boudoir photo, soft window light
  4  review   review ✓  real_person 0.99     Abraham Lincoln giving a speech, black and white photograph
  5  review   allow ✗   copyrighted_ip 0.99  Pikachu surfing a giant wave, anime style
  6  allow    allow ✓   copyrighted_ip 0.02  a bowl of ramen, studio lighting

Allowed 3, held 2 for review and blocked 1 of 6 prompts.
Matched the expected outcome on 5 of 6 (83.3%).

One prompt per line, or a CSV with a prompt column and an optional expected column. Up to 2,000 prompts a run, paced under the rate limit. The policy here is the family-app preset. TOP SCORE is the highest-scoring check; the policy's lines decide the outcome.

The command exits 1 when a prompt couldn't be checked or didn't get its expected outcome, as row 5 didn't. So a CI job can run a file of known prompts against a new policy version before you publish it: add --version 2 to a --dry-run.

Explain a decision

vindex explain dec_3ef40241be304ae0a23edb099e314615

REVIEW Sent to review by real_person (checks.real_person.on_act). Score 0.99 reached the act line 0.50, so checks.real_person.on_act applied: send to review.
dec_3ef40241be304ae0a23edb099e314615 · pol_cde75c4bb1f540e99bee051951dffa1e v1 · test key · 2026-09-17 10:20 UTC

Checks
  minors_sexual   score 0.01 · review score 0.01 · act line 0.50 · review line 0.30 · under its lines
    Score 0.01 and review score 0.01 are below the review line 0.30: nothing to do.
  adult           score 0.01 · review score 0.01 · act line 0.70 · review line 0.35 · under its lines
    Score 0.01 and review score 0.01 are below the review line 0.35: nothing to do.
  real_person     score 0.99 · review score 0.99 · act line 0.50 · review line 0.30 · over the act line: send to review
    Score 0.99 reached the act line 0.50, so checks.real_person.on_act applied: send to review.
    - a specific real, identifiable person: 0.99 against a line of 0.50.
  copyrighted_ip  score 0.04 · review score 0.04 · act line 0.50 · review line 0.30 · under its lines
    Score 0.04 and review score 0.04 are below the review line 0.30: nothing to do.

Actions
  Opened review item rev_000d6759762a465a90a9d9d0c99551be.

The same explanation as /explain, laid out for a terminal.

Read the decision log

vindex decisions --outcome review --since 1h

ID                                    OUTCOME  WHEN      LATENCY  USER  NOTES
dec_0e8c03890a1a4ae399f955108f862adb  review   9s ago     357 ms        test, dry run
dec_df14442cd9a44deb97e253112be1ff8e  review   9s ago     847 ms        test, dry run
dec_3ef40241be304ae0a23edb099e314615  review   18s ago    341 ms        test
dec_a537fadc78704d518caecc57e6c57800  review   5m ago   1,004 ms        test, dry run
dec_21cf7d5eb238415b91fdb8a961c69be3  review   13m ago  1,481 ms  u_42  test
dec_9d28719864c4453e98ac87fbc3a3e024  review   14m ago    401 ms        test, dry run
dec_b1aa487049cb4ee88328b756df8040a5  review   14m ago    820 ms        test, dry run

Newest first. --since takes 30m, 1h, 7d, 2w or a date. --limit goes up to 100, and a longer log prints the --cursor for the next page.

Rule on the review queue

vindex review
vindex review verdict rev_000d6759762a465a90a9d9d0c99551be allow --note "historical photo"

ID                                    OUTCOME  WAITING  TOP SCORE         USER  PROMPT
rev_000d6759762a465a90a9d9d0c99551be  review   10s ago  real_person 0.99        Abraham Lincoln giving a speech, black and white photograph

Rule on one: vindex review verdict <id> allow|block|dismiss
Recorded allow on rev_000d6759762a465a90a9d9d0c99551be (decision dec_3ef40241be304ae0a23edb099e314615).

Oldest first, like the review queue in the dashboard. A verdict is allow, block or dismiss, and an item takes one.

Every command

Every command takes --json, which prints the API's response as it came, and --help. An error prints the API's title and detail and exits 1. A rate-limited request waits once and tries again.

Connect an agent

The MCP server speaks Streamable HTTP at https://mcp.getvindex.com. Every request carries your API key as a bearer token, the same key you'd send to the API, and the server keeps no key of its own. Start with a test key.

claude mcp add --transport http vindex https://mcp.getvindex.com \
  --header "Authorization: Bearer $VINDEX_API_KEY"

{
  "mcpServers": {
    "vindex": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp.getvindex.com", "--header", "Authorization:${VINDEX_AUTH}"],
      "env": { "VINDEX_AUTH": "Bearer vx_test_..." }
    }
  }
}

{
  "mcpServers": {
    "vindex": {
      "url": "https://mcp.getvindex.com",
      "headers": { "Authorization": "Bearer ${env:VINDEX_API_KEY}" }
    }
  }
}

• Claude Code: the command saves the key in your Claude Code config. For a project's shared .mcp.json, write the header as "Authorization": "Bearer ${VINDEX_API_KEY}" and each person's own environment fills it in.
• Claude Desktop: paste into claude_desktop_config.json (Settings, Developer, Edit Config) and restart. Desktop's config runs local servers, so mcp-remote bridges to the remote one; it needs Node. Leave no space after Authorization: in args, because some clients split arguments on spaces; the space goes in VINDEX_AUTH.
• Cursor: ~/.cursor/mcp.json, or .cursor/mcp.json in a project. Cursor fills ${env:VINDEX_API_KEY} from your environment.

Without a key the server still connects and lists its tools, and every call answers with where to get one.

Tools

• check_content screens a prompt against your policy before anything is generated, and returns allow, review or block, why, and what to do. Takes prompt, and optionally policy, dry_run and user.
• explain_decision explains a past decision in plain words: which check and policy line decided it, each score against its line, and what ran. Takes id.
• list_policies lists your policies, their live versions and which is the default.
• get_policy says one policy's lines in plain words: at what score each check sends a prompt to review or acts, and what it does. Takes id.
• list_presets lists the presets a workspace can start from, one line each.
• bulk_check screens up to 50 prompts in one call and returns a table of outcomes, with a decision id per prompt. Takes prompts, and optionally dry_run and policy.
• review_queue lists open review items, oldest first. Optionally takes limit.
• review_verdict records a verdict on an open item: allow, block or dismiss. A verdict is final. Takes id and verdict, and optionally note.

check_content answers in two parts: a short one (the outcome, why, what to do, the decision id) and the decision's JSON.

The server also has a resource and a prompt:

• vindex://policy/{id}: a policy as text, its lines in words and then the document as JSON.
• screen_before_generating: a short instruction for an agent. Call check_content with the exact text before any image, video or text generation, then generate on allow, hold on review, and refuse on block without repeating the prompt.

Roles and limits are the API's. A key that can't rule on review items gets the API's error back, and every call counts against the key's rate limit.