DSPy

What it is for

DSPy is an MIT Python framework for programming language models: you declare a typed Signature, pick a Module (Predict / ChainOfThought / ReAct), then compile the program against a metric. Use it when the job is a maintainable LM program in a repo, not when you want a hosted chat seat with no code.

What makes it unique: the cover is the official stanfordnlp/dspy GitHub file list — dspy/, LICENSE (MIT), README.md, 3.4.1 — not a wordmark.

Key features

• Signatures — typed inputs and outputs instead of a prompt string (email -> event_name, date) (Program, don’t prompt; dspy.ai). The contract, not a generate.
• Modules — the same signature can run as Predict, ChainOfThought, or ReAct (dspy.ai “Same interface, different strategy”). An execution strategy. Tool lists you pass into ReAct are this act stream, not a second product.
• Optimizers — GEPA / MIPROv2 compile prompts and demos against a metric you write (optimizer.compile(program, trainset)) (dspy.ai; Jul 2025 GEPA paper). A compile job, not a chat.
• PythonInterpreter — a sandboxed backend that runs code the program wrote (3.4.1 notes; homepage calc tool). Execution, not another prompt.
• Multimodal fields — dspy.Image is a Signature field type so a chart or photo is an input, not a caption you pasted (dspy.ai multimodal tab). A type, not a module.

pip install -U dspy and dspy.LM("openai/…") strings are install / SDK. Model names are SKU. You bring the provider bill.

Who it's for

Best fit: Python teams who will write signatures and a metric and already have a model API key.

Pricing

Verified 2026-10-22 (Asia/Shanghai) from the MIT LICENSE on the repo cover and dspy.ai (pip install -U dspy). There is no hosted SKU table on dspy.ai.

• Software — MIT. License fee $0.
• Model / hosting bill — you pay the provider you wire. DSPy does not print those dollars.

Do not invent a DSPy Cloud seat.

Limitations / costs

• You bring the model bill. The library is $0; tokens are not.
• Code-first. There is no dspy.ai seat table. For a hosted TypeScript agent platform, see Mastra.
• Library, not a paper writer. For cited academic drafts, see Yomu.

Alternatives

• Pydantic AI — Python agent SDK with a public license.
• Mastra — TypeScript agents + workflows on Mastra Platform.
• Haystack — RAG/agent pipelines with Studio $0 / Enterprise custom.

Official tutorial

We searched for a first-party DSPy / StanfordNLP YouTube getting-started lesson we could oEmbed. Conference or Cohere interviews (including Omar Khattab on @CohereAI) are not embedded. Use written Installation (pip install dspy / uv add dspy) plus Program, don’t prompt.

Website

https://dspy.ai

Repository

https://github.com/stanfordnlp/dspy

Source

Catalog row curated for AI Tools Nav. Cover is the official stanfordnlp/dspy GitHub file list (MIT). License fee $0; no dspy.ai seat table. No first-party YouTube embed.

Last verified

2026-10-22 (Asia/Shanghai) editorial check.