Perspective API Alternative: Toxicity Detection API & MCP Server

Google stops the Perspective API on December 31, 2026. Google gives no migration support. This Actor is a direct replacement. It uses the same request format, the same response format and the same attribute names. To move to this Actor, change the URL only.

• Perspective-compatible endpoint: POST /v1alpha1/comments:analyze
• Six attributes: TOXICITY, SEVERE_TOXICITY, INSULT, PROFANITY, THREAT, IDENTITY_ATTACK
• Batch endpoint: send a maximum of 100 comments in one call. A Console run accepts a maximum of 10,000 comments.
• MCP server: AI agents (Claude, Cursor and other agents) can examine text before they publish it.
• No LLM: the model is fast and deterministic. The same text always gets the same score.
• No text storage through the API: the server keeps API and MCP requests in memory only. It does not store or log them. A Console run saves the results to your own dataset.
• Long comments: the model divides a long text into chunks that overlap. Thus, it examines the end of the text too. In a very long, polite text, one insult can get a score below 0.5. For documents, use spanAnnotations or a lower threshold.

Change to this Actor in 1 minute

- https://commentanalyzer.googleapis.com/v1alpha1/comments:analyze?key=GOOGLE_API_KEY
+ https://koyourmoon--perspective-api-alternative.apify.actor/v1alpha1/comments:analyze?token=APIFY_TOKEN

Do not change the request body:

import requests

URL = "https://koyourmoon--perspective-api-alternative.apify.actor/v1alpha1/comments:analyze"
body = {
    "comment": {"text": "What kind of idiot name is foo?"},
    "languages": ["en"],
    "requestedAttributes": {"TOXICITY": {}, "INSULT": {}},
}
r = requests.post(URL, json=body, headers={"Authorization": "Bearer <APIFY_TOKEN>"})
print(r.json()["attributeScores"]["TOXICITY"]["summaryScore"]["value"])  # 0.93

The response has the same format as the Perspective response:

{
  "attributeScores": {
    "TOXICITY": {
      "spanScores": [{"begin": 0, "end": 31, "score": {"value": 0.926868, "type": "PROBABILITY"}}],
      "summaryScore": {"value": 0.926868, "type": "PROBABILITY"}
    }
  },
  "languages": ["en"],
  "detectedLanguages": ["en"]
}

The endpoint accepts these request options: spanAnnotations (scores for each sentence), scoreThreshold, clientToken, languages and doNotStore.

The endpoint also accepts attribute names with the _EXPERIMENTAL suffix (for example, TOXICITY_EXPERIMENTAL). These names are aliases.

Error responses use the Perspective errorType values: ATTRIBUTE_NOT_SUPPORTED, LANGUAGE_NOT_SUPPORTED_BY_ATTRIBUTE, COMMENT_EMPTY and COMMENT_TOO_LONG.

Coral by Vox Media (Talk): change two settings

Coral sends each new comment to the Perspective API. You can send these comments to this Actor. Coral keeps the query of the endpoint URL. Thus, you can put your Apify token in the endpoint.

1. In the Coral admin, open Configure > Moderation.
2. Find the settings of the toxic comment filter (Perspective).
3. In the endpoint field, enter https://koyourmoon--perspective-api-alternative.apify.actor/v1alpha1?token=<APIFY_TOKEN>.
4. In the key field, enter any text. The Actor does not use this key.
5. Keep your model (TOXICITY or SEVERE_TOXICITY) and your threshold.
6. Save the settings.

In the Coral GraphQL API, these fields are endpoint and key of the Perspective integration.

CAUTION: Increase the Coral timeout before you use this endpoint. The default Coral timeout for Perspective calls is 800 ms (environment variable PERSPECTIVE_TIMEOUT). If a call takes more time, Coral publishes the comment without a toxicity score. We recommend PERSPECTIVE_TIMEOUT=3000. After an idle period of 60 seconds, the first call can take 10 to 30 seconds (cold start). Thus, on a site with little traffic, Coral can publish some comments without a score.

The Actor also accepts the Perspective feedback calls (comments:suggestscore). It does not store them, and it does not charge for them.

Use the same method for other clients that let you change the Perspective endpoint.

Batch endpoint

curl -X POST "https://koyourmoon--perspective-api-alternative.apify.actor/v1/analyze" \
  -H "Authorization: Bearer $APIFY_TOKEN" -H "Content-Type: application/json" \
  -d '{"texts": ["Great post!", "You are an idiot"], "attributes": ["TOXICITY","INSULT"], "threshold": 0.5}'

{"results": [
  {"text": "Great post!", "scores": {"TOXICITY": 0.0006, "INSULT": 0.0002}, "flagged": false, "billed_units": 1},
  {"text": "You are an idiot", "scores": {"TOXICITY": 0.98, "INSULT": 0.95}, "flagged": true, "billed_units": 1}
]}

MCP server for AI agents

Add this text to the configuration of your MCP client:

{
  "mcpServers": {
    "toxicity": {
      "url": "https://koyourmoon--perspective-api-alternative.apify.actor/mcp",
      "headers": {"Authorization": "Bearer <APIFY_TOKEN>"}
    }
  }
}

The server has one tool: analyze_toxicity(texts, attributes?, threshold?). Use this tool to examine comments, replies or LLM output before your agent publishes them.

Runs without code

1. Open the Actor in the Console.
2. Paste a maximum of 10,000 texts. Put one text on each line.
3. Start the run.
4. Download the scores as CSV, JSON or Excel.

Prices

Event	Price
Text unit (each started 1,000 characters of a comment)	$0.50 for 1,000 units ($0.0005 for each unit)
Actor start	$0.002 for each GB of memory (about $0.004 for each start at the default 2 GB)

Most comments have less than 1,000 characters. Thus, most comments cost 1 unit ($0.0005). A longer comment costs more units. For example, a comment with 2,500 characters costs 3 units. Azure Content Moderator uses the same method: it charges for each 1,000 characters.

Batch results and Console results show billed_units. There is no subscription and no minimum. You do not pay for requests that are not valid.

• Console runs and API runs: you pay only the prices in the table.
• Standby API and MCP: while the server operates, Apify charges the platform usage of the server to your account. Apify does this for all Standby Actors.
• The server stops after 60 seconds without requests. Thus, a small quantity of requests costs very little.
• For a constant flow of requests, send the comments in batches of a maximum of 100 with POST /v1/analyze.

Comparison with Perspective

	Perspective API	This Actor
Availability	Stops on December 31, 2026	Available
Request and response format	v1alpha1	Same
Attributes	6 production + experimental	6 production (+ _EXPERIMENTAL aliases)
Languages	Many	English (more languages in the plan)
Text storage	Optional	The API and MCP do not store text

About the scores. The model is the open-source Detoxify "original" model. Jigsaw, the team that made Perspective, released the Toxic Comment dataset. Detoxify learned from this dataset.

The scores tell almost the same thing as the Perspective scores. But the values are not the same. If you use fixed thresholds, compare them with a sample of your own data after you change to this Actor.

Frequently asked questions

Does this Actor have a relation with Google or Jigsaw? No. This Actor is an independent, compatible service.

Can I use it for text that is not in English? Version 1 supports English only. The endpoint returns LANGUAGE_NOT_SUPPORTED_BY_ATTRIBUTE in these conditions:

• The request has a different value in languages.
• The request has no languages value, and the text is clearly in a non-Latin script (for example, Korean, Cyrillic, Chinese or Arabic).

Perspective returns the same error for languages that it does not support. Thus, you do not pay for a score that you cannot use. The batch endpoint gives scores for these texts, but it adds a warning field to the result.

How fast is it? When the server operates, a short comment takes tens of milliseconds. After a period without requests, the first request starts the server. This cold start usually takes 10 to 30 seconds. Set a client timeout of a minimum of 60 seconds. Standby mode adds servers automatically when the load increases.

What is the maximum length of a comment? The maximum length is 20,480 bytes. Perspective has the same limit.