From idea to production in just few lines

The first neuro-symbolic Language Model (LM) framework leveraging the simplicity of Keras and the rigor of Deep Learning best practices.

Build RAGs, tool-using agents, multi-agents systems, recursive agents and more in just few lines

Deutsch | English | Español | Français | 日本語 | 한국어 | Português | Русский | 中文

Documentation · FAQ · Discord · Code Examples . Guides

If you find Synalinks useful, please star the repo! Help us reach more AI/ML engineers and grow the community.

Want to use Synalinks with your own coding agent (Claude Code, Cursor, Copilot, etc.)? Add the Synalinks-specific skills from synalinks-skills on GitHub to your agent — they teach it the framework conventions and give it the context it needs to build Synalinks programs right away.

What Is Synalinks?

Synalinks is an open-source neuro-symbolic framework that makes it simple to create, train, evaluate, and deploy advanced LM-based applications, including RAGs, autonomous agents, and self-evolving reasoning systems.

Think Keras for Language Models applications, a clean, declarative API where:

• You compose Modules like you would with deep learning Layers.
• You train & optimize with in-context reinforcement learning.
• You deploy as REST APIs or MCP servers.

Key Principles

• Progressive complexity: Start simple and grow advanced naturally.
• Neuro-symbolic learning: Combine logic, structure, and language models.
• In-context optimization: Improve model reasoning without retraining weights.

Who Is It For?

Role	Why Synalinks Helps
AI Developers	Build complex production grade LM apps without boilerplate.
AI Researchers	Prototype neuro-symbolic and RL-in-context systems fast.
Data Scientists	Integrate LM workflows with APIs & databases.
Students/Hobbyists	Learn AI composition in a clean, intuitive framework.

Why Synalinks?

Many frameworks exist today — here is what Synalinks does differently:

• Embedded, container-free sandbox : agents run untrusted code and tools in a safe, isolated runtime that needs no Docker or external sandbox service. The whole stack is pure-Python and embeddable, so it is great for scripting, research, serverless/cloud deployment (S3, Lambda, notebooks, etc.) or even for creating CLI harnesses!
• Embedded database support : build graph-based RAG and agentic memories with constrained Knowledge Graph extraction and automatic semantic deduplication, on top of an embedded graph database — no separate graph server to run. Additionally, a fast embedded SQL knowledge base is available to store relational data and build vector/SQL RAGs.
• In-Context RL to optimize your prompts (and anything else) : train and optimize prompts, few-shot examples, and any trainable variable per module without touching model weights, using the familiar .compile() / .fit() / .evaluate() / .predict() API.
• Effortless model switching : set a default once with synalinks.set_default_language_model(...) or pass a string identifier, and swap between Ollama, vLLM, OpenAI, Azure, Anthropic, Mistral, Groq, Gemini, xAI, Cohere, DeepSeek, Together AI, OpenRouter, AWS Bedrock and Doubleword via LiteLLM — including multi-objective model selection to pick the best model for cost/quality.
• Scaffold in one command, bring your own coding agent : bootstrap a production-ready project with synalinks init (batteries-included templates for scripts, agents, and training), then drop in the official Synalinks skills so Claude Code, Cursor, Copilot and friends write idiomatic Synalinks code from the start.

Plus everything you'd expect from a production-grade framework:

• NEW: Now all agents support Agent Skills, AGENTS.md and sub-agents.
• Constrained structured outputs (JSON) for correctness
• Chat Completions-compatible message API: messages mirror the OpenAI Chat Completions format key-for-key + litellm-extended with reasoning_content and thinking_blocks so provider reasoning survives multi-turn round-trips.Also handle multimodal inputs (images & audio as standard content parts)
• Versionable, JSON-serializable pipelines
• Automatic async & parallel execution by default
• Metrics, rewards & datasets built-in
• API-ready: Deploy with FastAPI or FastMCP
• KerasTuner compatibility for hyperparameter search
• Built-in callbacks and hooks for observability (including an MLflow Monitor callback)

Requirements

• Python 3.12 or more
• WSL2 for windows users

Quickstart in 3s with uv (recommended)

If you don't know uv, install it here.

Follow the instructions to start a new synalinks project in 3s:

uvx synalinks init

---

You can also install the library in a new project with:

uv add synalinks

To transform your coding agent into an AI engineer, do this at the root of your project:

npx skills add SynaLinks/synalinks-skills --skill synalinks

Example

Synalinks agents can now read your project's AGENTS.md conventions and use Agent Skills. The example below wires the official Synalinks skills into a DeepAgent — a sandboxed coding agent — and asks it to design the input/output data models for a task, writing them into a workspace/ folder.

First set up the workspace. Install the official Synalinks skill with the skills CLI and add an AGENTS.md. The skill lands under the workdir, so the sandboxed agent can read its body on demand:

mkdir -p workspace && cd workspace
# Installs the `synalinks` skill into ./.agents/skills/ and writes skills-lock.json.
npx skills add SynaLinks/synalinks-skills --skill synalinks

This gives the layout below — .agents/skills is the skills root (one sub-folder per skill, each holding a SKILL.md):

workspace/
├── AGENTS.md                     # injected as the agent's conventions
├── skills-lock.json              # pins the skill to a source repo + content hash
└── .agents/
    └── skills/                   # the skills root
        └── synalinks/
            └── SKILL.md          # name + description surfaced; body read on demand

main.py:

import synalinks
import asyncio

# Set the default once — modules pick it up automatically.
synalinks.set_default_language_model("gemini/gemini-3.1-flash-lite-preview")

# The agent's structured final answer.
class Deliverable(synalinks.DataModel):
    summary: str = synalinks.Field(
        description="What was created and where",
    )
    files: list[str] = synalinks.Field(
        description="Paths of the files written into the workspace",
    )

async def main():
    # A DeepAgent converses in ChatMessages (it is a coding agent).
    inputs = synalinks.Input(data_model=synalinks.ChatMessages)

    agent = synalinks.DeepAgent(
        data_model=Deliverable,
        # The sandbox is seeded from this directory (host-safe: the agent's
        # writes land in the sandbox copy, never on your disk). Its `AGENTS.md` is
        # injected so the agent follows your conventions.
        workdir="workspace",
        # The skills root (installed by `skills add`). Listed to the agent as
        # `<available_skills>`; it reads each `SKILL.md` on demand from the
        # sandbox — which is why the skills live under `workdir`.
        skills=["workspace/.agents/skills"],
    )
    outputs = await agent(inputs)

    program = synalinks.Program(
        inputs=inputs,
        outputs=outputs,
        name="datamodel_designer",
        description="Designs Synalinks data models for a given task",
    )

    task = (
        "Define the input and output Synalinks DataModels for a support-ticket "
        "triage task: the input is a raw customer message; the output is the "
        "predicted category, a priority, and a short suggested reply. Write them "
        "to `models.py` using idiomatic Synalinks — consult the skills first."
    )
    result = await program(
        synalinks.ChatMessages(
            messages=[synalinks.ChatMessage(role="user", content=task)],
        )
    )
    print(result.prettify_json())

if __name__ == "__main__":
    asyncio.run(main())

Data Model Operators

Synalinks provides Python operators for combining and manipulating data models, enabling sophisticated control flow. See the Control Flow guide for the routing, fan-out, and merge patterns these operators enable:

Operator	Name	Description	Use Case
+	Concatenation	Combines fields from both data models. Raises exception if either is None.	Merging outputs from parallel branches
&	Logical And	Safe concatenation that returns None if either input is None.	Combining with potentially null branch outputs
|	Logical Or	Returns the non-None data model. If both are non-None, merges them.	Gathering outputs from conditional branches
^	Logical Xor	Returns data if exactly one input is non-None, otherwise None.	Exclusive branch selection
~	Logical Not	Returns None if input is non-None, or a empty data model if None.	Inverting branch conditions
in	Contains	Checks if a string key exists in the schema properties, or if another data model's schema is contained. Returns True or False.	Conditional field checking, schema validation

# Parallel branches with concatenation
x1 = await generator1(inputs)
x2 = await generator2(inputs)
# combined = x1 *and* x2
combined = x1 & x2  # Merge both outputs (add _{i} suffix if key collision)
# [...]
# Conditional branches with logical or
(easy, hard) = await synalinks.Branch(
    question="Is this query complex?",
    labels=["easy", "hard"],
    branches=[simple_generator, complex_generator],
)(inputs)
# result = easy *or* hard
result = easy | hard  # Get whichever branch was selected

Getting a summary of your program

To print a tabular summary of your program:

program.summary()

Or a plot (Useful to document your system):

synalinks.utils.plot_program(
    program,
    show_module_names=True,
    show_trainable=True,
    show_schemas=True,
)

The data model designer program visualized with plot_program: Input → DeepAgent. Trainable modules are marked in green.

Running your program

To run your program use the following:

result = await program(
    Query(
        query=(
            "A bookstore receives a shipment of 135 new books."
            "They place the books evenly onto 9 shelves."
            "Later, they decide to move 3 books from each shelf to a display table"
            " at the front of the store. "
            "How many books are left on the shelves after the books are moved?"
        )
    ),
)

Training your program/agent

# Setting the default language/embedding models allows you
# to use the string identifier (Keras-like) to configure your pipeline/training.
# You can still instantiate the classes if you want fine-grained control.
synalinks.set_default_language_model("gemini/gemini-3.1-flash-lite-preview")
synalinks.set_default_embedding_model("gemini/text-embedding-004")

async def main():
    

    # ... your program definition

    (x_train, y_train), (x_test, y_test) = synalinks.datasets.gsm8k.load_data()

    program.compile(
        reward=synalinks.rewards.ExactMatch(in_mask=["answer"]),
        optimizer="omega",
    )

    batch_size=1
    epochs=10

    history = await program.fit(
        x_train,
        y_train,
        validation_split=0.2,
        batch_size=batch_size,
        epochs=epochs,
    )

if __name__ == "__main__":
    asyncio.run(main())

Saving & Loading

To save the entire architecture and variables (the program's state) into a JSON file, do:

program.save("my_program.json")

In order to load it, do:

loaded_program = synalinks.Program.load("my_program.json")

To save only the state your program (the variables) into JSON:

program.save_variables("my_program.variables.json")

To load its variables (needs a program with the same architecture), do:

program.load_variables("my_program.variables.json")

Logging

To enable logging, use the following at the beginning of your script:

synalinks.enable_logging()

Observability

Synalinks provides built-in observability through MLflow for tracing and monitoring your programs.

Important: Call enable_observability() before creating any modules.

import synalinks

# Enable observability first
synalinks.enable_observability(
    tracking_uri="http://localhost:5000",  # Optional: MLflow server URI
    experiment_name="my_experiment"         # Optional: defaults to "synalinks_traces"
)

# Then create your modules - they will be automatically traced
inputs = synalinks.Input(data_model=Query)
outputs = await synalinks.Generator(...)(inputs)

For training metrics and artifacts, use the Monitor callback:

monitor = synalinks.callbacks.Monitor(
    tracking_uri="http://localhost:5000",
    experiment_name="training_runs",
)

await program.fit(x=train_x, y=train_y, callbacks=[monitor])

See the Observability guide for advanced configuration.

Learn more

You can learn more by reading our documentation. If you have questions, the FAQ might help you.

Contributions

Contributions are welcome, either for the implementation of additional modules, metrics, or optimizers. For more information, or help for implementing your ideas (or ones from a paper), please join our discord.

Beware that every additional metric/module/optimizer should be approved by the core team, we want to keep the library minimal and clean as possible to avoid an uncontrolled growth leading to bad software practices like in most current leading LM frameworks.

If you have specific feedbacks or features request we invite you to open an issue.

Contributors

Your contributions, feedback, and support are what make this project thrive.

From small bug fixes to major features, thank you for believing in open collaboration and the future of neuro-symbolic AI.

Community

Join our community to learn more about neuro-symbolic systems and the future of AI. We welcome the participation of people from very different backgrounds or education levels.

Citing our work

This work have been done under the supervision of François Chollet, the author of Keras. If this work is useful for your research please use the following bibtex entry:

@misc{sallami2025synalinks,
  title={Synalinks},
  author={Sallami, Yoan and Chollet, Fran\c{c}ois},
  year={2025},
  howpublished={\url{https://github.com/SynaLinks/Synalinks}},
}

Credit

Synalinks would not be possible without the great work of the following open-source projects:

• Keras for the graph-based computation backbone, API and overall code, design and philosophy.
• DSPy for the modules/optimizers inspiration.
• Pydantic for the backend data layer.
• LiteLLM for the LMs integrations.
• DuckDB, Ladybug, LanceDB for their amazing embedded databases.
• MirageAI for their amazing sandbox!