uClassify
AIScore text against thousands of public classifiers with the uClassify REST API, covering sentiment, topic, and more. Agents route and prioritize inbound text without training a model first.
What This Integration Enables
The usual objection to classifying inbound text is that you have to build the classifier first, which means labelling data, which means the project dies. uClassify removes that step for the common cases. Thousands of public classifiers already exist in its directory, covering sentiment, topic, language, gender, ageing, and the IAB taxonomy, and an agent can call any of them by name with no training whatsoever. You find out whether classification helps your routing before you invest in making it good.
Two details make the output usable inside a flow rather than just interesting. Classify Text returns the full probability distribution across every class, not a single winning label, so a flow can branch on the margin between the top two candidates instead of pretending a coin flip was a decision. And it returns a text coverage score reporting how much of the input the classifier actually recognised, which is the honest signal that a classifier trained on product reviews is being handed a legal notice. When either number is weak, that is not a classification. That is an automation exception, and it should go to a person. Extract Keywords is the explainability companion: it shows exactly which words pushed a review negative, which is what a human needs in order to overrule the machine with confidence. When the generic classifiers do run out, Create Classifier, Add Class, and Train Classifier let you build your own on the same API.
Without FlowRunner
With FlowRunner
Use Case Scenarios
Sentiment escalation out of a shared inbox
Every message arriving in Gmail is scored for sentiment. Neutral and positive messages route normally. Strongly negative messages are escalated into a Slack channel with the message, the probability, and the keywords Extract Keywords identified as the drivers, so the person picking it up knows in one glance whether this is a churn risk or somebody's shift key. Messages where the negative and neutral probabilities sit close together do not get escalated and do not get buried. They go to a triage view, because a message the classifier is unsure about is exactly the one a human should read.
Content categorization at publish time
A post is published in WordPress. The agent classifies the body against the IAB taxonomy classifier and writes the resulting categories, with their probabilities, into a content index in Airtable. Advertising, content recommendations, and internal reporting all read from the same index rather than each maintaining their own tagging. Where the text coverage score comes back low, usually because the post is mostly code samples or a table, the row is flagged for manual tagging instead of being given a confident-looking category that is wrong.
A classifier that learns from the queue it feeds
Public classifiers get you started. They do not know your product names or your escalation categories. So the agent builds one: Create Classifier, then Add Class for each routing destination, then Train Classifier with examples drawn from tickets that were routed and resolved. From then on the flow calls the custom classifier alongside the public sentiment one, and every correction a triager makes becomes a candidate training example. What was a static rule set becomes a routing model that reflects how the team actually works, without a data science project attached to it.
Human-in-Loop Highlight
The tempting design here is a closed loop: a triager corrects a misrouted ticket, the correction goes straight into Train Classifier, the classifier improves. It is also the design that will quietly ruin the classifier. Training on uClassify is incremental and cumulative, so every example fed to a class stays in the model permanently, and the only way to reverse a bad example is Untrain Classifier with exactly the original text that was trained. Not a paraphrase. The exact string. If a batch job trained a hundred examples from a misconfigured field mapping, and nobody kept the strings, that damage is not recoverable. Remove Class and Remove Classifier discard training data outright and cannot be undone at all.
So FlowRunner agents route freely and train under supervision. The routing side is fully automatic: classify, check the margin and coverage, send it where it belongs. The learning side stops. When corrections accumulate, the agent batches them and puts them in front of the person who owns the classifier: "Fourteen tickets routed to Billing were reassigned to Provisioning this week. Here are the fourteen texts and the keywords that pushed each one to Billing. Train these into Provisioning, train a subset, or leave the classifier alone because these were genuinely ambiguous?" The agent keeps the exact strings it trained so an Untrain is possible later, and no example enters the model that a person did not agree to. Human-in-the-loop sits at the write that cannot be walked back, not at the read that happens ten thousand times a day.
Agent Capabilities
9 actionsClassification
3- Classify Text Sends one or more texts to a uClassify classifier and returns, for each text, the probability of every class along with a text coverage score indicating how much of the input the classifier recognised. Works with the thousands of public classifiers in the uClassify directory, covering sentiment, topics, language, ageing, gender, and the IAB taxonomy, as well as classifiers you have trained yourself. Requests are capped at 3 MB.
- Extract Keywords Sends one or more texts to a uClassify classifier and returns the individual keywords that drove the classification, each with the class it points to and its probability. The explainability companion to Classify Text: it shows exactly which words made a review read as negative or a page read as a given topic.
- Get Classifier Information Retrieves the classes defined on a classifier together with the size of each class's trained model, reported as the number of unique features and the total feature count. Used to confirm a classifier's class names before wiring up a flow, or to monitor how much training data a classifier you own has accumulated.
Classifier Management
4- Create Classifier Creates a new, empty classifier on your uClassify account. Classes are added to it and trained before it can classify anything. Requires the Write API Key.
- Remove Classifier Permanently deletes a classifier from your uClassify account, including every class and all of its accumulated training data. This cannot be undone. Requires the Write API Key.
- Add Class Adds a new class to a classifier you own. A classifier needs at least two classes to produce a meaningful probability distribution, and each class must then be trained with representative example texts. Requires the Write API Key.
- Remove Class Deletes a class from a classifier you own, discarding all training data associated with that class while leaving the rest of the classifier intact. Requires the Write API Key.
Training
2- Train Classifier Teaches a class of one of your classifiers by feeding it example texts that belong to that class. Training is incremental, so this can be called repeatedly as new labelled examples arrive, and accuracy improves as each class accumulates more representative material. Requests are capped at 3 MB. Requires the Write API Key.
- Untrain Classifier Removes previously supplied example texts from a class, reversing their effect on the trained model. Used to correct mislabelled training data without rebuilding the classifier from scratch. Requires exactly the texts that were originally trained. Requires the Write API Key.
Frequently Asked Questions
What can FlowRunner do with uClassify?
FlowRunner agents can run Classify Text, Extract Keywords, and Get Classifier Information in uClassify, plus 6 more actions.
Does connecting uClassify to FlowRunner require OAuth?
No. uClassify connects to FlowRunner with an API key, no OAuth flow required.
Can uClassify trigger a FlowRunner workflow automatically?
uClassify doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.
Start building with uClassify
$100 in credits. No card required. Connect in minutes.