DeepSeek

Run DeepSeek large language models with extended reasoning, tool and function calling, JSON output, and prefix completion. Fast, low-cost inference with an OpenAI-compatible interface.

6 actions available

New batch of transactions flagged for review

Agent sends each transaction to DeepSeek in reasoning mode

Model returns a risk assessment with its reasoning

Agent clears low-risk transactions by rule

Agent records the assessment against each transaction

Analyst gets a queue of the flagged transactions

High-risk or borderline transactions route to an analyst to decide

What This Integration Enables

DeepSeek pairs strong reasoning with low cost, which changes what is economical to automate. Its thinking mode works through a problem step by step, tool calling lets it invoke functions mid-task, and JSON output makes the result machine-readable. Prefix completion gives fine control over how a generation starts. Cheap reasoning means you can run a model first-pass on volumes that were previously too expensive to bother with. But a first pass is a first pass. The point is to have the model do the legwork and surface its reasoning so a person can act on it faster. An orchestration layer routes the model's assessments and pulls the analyst in on the ones that matter, and FlowRunner is built for that layer.

Without FlowRunner

Reasoning costs add up Step-by-step tasks run on expensive frontier models

Manual first-pass review An analyst reads every flagged item from scratch

Opaque model output A verdict with no visible reasoning to check

With FlowRunner

Low-cost reasoning Extended reasoning at a fraction of frontier pricing

Pre-assessed queue The model does the first pass, the analyst reviews

Reasoning visible The model shows its work so a person can check it

Use Case Scenarios

Transaction Review First-Pass

A batch of transactions is flagged for review. The agent sends each to DeepSeek in reasoning mode and gets back a risk assessment with the reasoning behind it. Low-risk items clear by rule. The analyst opens a pre-assessed queue and reviews the model's reasoning instead of starting from scratch, and the borderline cases are the ones that reach a person.

Bulk Content Classification

A large volume of content needs categorizing on a budget. The agent classifies each item with DeepSeek, using JSON output so the results drop straight into the system. The low cost per call makes it economical to classify everything rather than a sample, and items the model is unsure about are set aside for a person.

Structured Data Extraction

Documents need specific fields pulled out reliably. The agent prompts DeepSeek with a schema and gets structured JSON back. Records that validate flow through automatically. Anything that fails validation or comes back with low confidence is routed to a person, so the pipeline stays clean without a frontier-model bill.

Human-in-Loop Highlight

A model's risk assessment is a recommendation with reasoning attached, not a ruling. When DeepSeek flags a transaction as high-risk or returns a borderline assessment, FlowRunner routes it through a [human-in-the-loop](/concepts/human-in-the-loop/) step: the agent pauses, presents the item along with the model's reasoning, and sends it to an analyst via Slack. They read the reasoning and decide. Cheap reasoning lets the model assess everything; a person owns the calls with consequences.

Agent processes routinely

Detects exception requiring judgment

Clear match Continues automatically

Ambiguous Routes to human via Slack

Human decides

Agent resumes with decision

Agent Capabilities

6 actions

Chat

Chat Completion Generates a text response for a single prompt using a DeepSeek model (deepseek-v4-flash or deepseek-v4-pro, both with a 1M-token context window and up to 384K output tokens).
Chat Completion (Advanced) Sends a fully custom chat completion request to DeepSeek with a complete messages array (multi-turn conversations with system, user, assistant and tool roles), tool/function calling passthrough (up to 128 functions), structured output via a response format object, thinking mode and reasoning effort controls, log probabilities, and sampling parameters.
Chat Prefix Completion Completes a response that must start with the exact text you provide, using DeepSeek's beta chat prefix completion feature.

Completions

FIM Completion Generates a fill-in-the-middle (FIM) completion using DeepSeek's beta completions endpoint, ideal for code completion where the text before (prompt) and after (suffix) the insertion point are known.

Models

List Models Lists all models currently available through the DeepSeek API (e.g. deepseek-v4-flash and deepseek-v4-pro) with their IDs and owning organization.

Account

Get Balance Retrieves the DeepSeek account's current balance, including whether the balance is sufficient for API calls and a per-currency (USD/CNY) breakdown of total, granted and topped-up balances.

Start building with DeepSeek

$100 in credits. No card required. Connect in minutes.

Start Building Free Talk to us