Sentiment Analysis for AI Agents

Overview

Sentiment Analysis runs two independent scorers over the `text` you POST. The first is lexicon-based: the text is tokenised, each token is looked up in the AFINN list of integer valences from -5 to +5, and the matches are summed into `sentiment`. `comparative` divides that sum by the token count. `sentimentText` maps the sum to a label with fixed cut-offs: 4 and above is very positive, 2 to 3 positive, -1 to 1 neutral, -2 negative, and -3 or below very negative. `isPositive` and `isNegative` are only the sign of the sum, so a score of 1 returns `isPositive: true` while the label still reads neutral.

On paid plans the text is also split into sentences at terminal punctuation and line breaks, and the first 50 sentences go to a transformer classifier trained on social media text. Each sentence gets positive, neutral and negative probabilities. `modelScore` is the mean positive probability minus the mean negative probability, rounded to two decimals, and `modelSentiment` is the label with the highest mean. `isMixed` is true when at least one sentence is labelled positive and another negative, each with confidence of 0.6 or more; `modelSentiment` then reads `mixed`.

`normalizedScore` is the AFINN sum divided by 10 and clamped to -1..1, so any text scoring 10 or more in either direction saturates. `magnitude` is the sum of absolute valences and grows with length: use it to tell strongly worded text from flat text, not to compare documents of different sizes. The free plan skips the model call entirely, so the model fields are null there.

Live Test Sentiment Analysis for AI Agents Source →

The tool

Once your client is connected to the VerveContext server, this appears in its tool list as SentimentAnalysisforAIAgents. It is read-only and open-world — it fetches and never mutates anything on your side — so most clients call it without asking you to confirm.

{
  "name": "SentimentAnalysisforAIAgents",
  "arguments": {
    "text": "I absolutely love this product! It exceeded all my expectations."
  }
}

You do not name the tool yourself; the model picks it. Asking about I absolutely love this product! It exceeded all my expectations. in the terms this source covers is enough for it to reach for SentimentAnalysisforAIAgents on its own — naming it explicitly also works, and is the way to force the call.

Connecting

One server URL covers every source in the catalog, including this one. Authorization is OAuth: the client opens a browser once, and there is no key to paste into a config file.

{
  "mcpServers": {
    "vervecontext": {
      "url": "https://api.vervecontext.com/v1/mcp"
    }
  }
}

https://api.vervecontext.com/v1/mcp

Per-client setup — Claude, Cursor, VS Code, ChatGPT — is on the MCP setup page.

Arguments

These are the properties on the tool's inputSchema, so a well-behaved client validates them before the call is made. Premium arguments are accepted on every plan but only take effect on plans that include them.

What the model gets back

The result carries a structuredContent object matching the tool's declared outputSchema, so a client reads fields without parsing prose. status is "ok" and error is null on success; a null field means the value was not available for that input, not that the call failed.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42

{
  "status": "ok",
  "error": null,
  "data": {
    "comparative": 0.25,
    "sentimentText": "positive",
    "sentiment": 3,
    "isPositive": true,
    "isNegative": false,
    "normalizedScore": 0.3,
    "magnitude": 3,
    "positiveWords": [
      "excited"
    ],
    "negativeWords": [],
    "modelSentiment": "positive",
    "modelScore": 0.98,
    "isMixed": false,
    "sentences": [
      {
        "text": "I'm so excited that tomorrow is going to be sunny!",
        "label": "positive",
        "confidence": 0.992,
        "scores": {
          "positive": 0.992,
          "neutral": 0.006,
          "negative": 0.002
        }
      },
      {
        "text": "Can't wait!",
        "label": "positive",
        "confidence": 0.973,
        "scores": {
          "positive": 0.973,
          "neutral": 0.024,
          "negative": 0.003
        }
      }
    ]
  }
}

Response fields

Paths are relative to data. Premium fields are absent rather than zeroed on plans that do not include them, so check for presence instead of comparing to 0.

Why ground on it

A model can produce something that looks like this answer from its training data, and be confidently out of date or simply wrong. This source returns the current value in a shape you can check, which is the difference between an answer you can cite and one you have to hedge.

Point an evaluation at comparative: it is the field most worth pinning a claim to, and it is either present and current or absent — never plausibly invented.

Failure modes

Errors come back as tool errors carrying a sentence the model can act on, not a bare status code. Error handling covers the full list.

A call costs 10 credits each time the tool actually runs; a model that reasons about the tool without calling it costs nothing.

Use cases

Support Ticket Escalation

Route angry customer messages to senior support reps immediately when the sentiment score drops to very negative.

Review Stream Filtering

Product managers categorize incoming app store feedback into positive praise and negative complaints using the five-tier sentiment label.

Social Media Sentiment Analysis

When tracking brand tags on social networks, marketing teams flag sudden spikes in negative valence scores to address PR issues early.

Survey Response Sorting

Sort open-ended employee survey responses automatically by comparative sentiment to isolate recurring workplace grievances without manual reading.

Other ways to use Sentiment Analysis for AI Agents

Set up Sentiment Analysis for AI Agents on VerveContext, or reach the same source a different way. Your VerveContext account and credits work on all of them — one key, one balance.

Related

More in AI/Computer Vision: