An open-source Deep Agent assistant for biologists, bioinformaticians, and clinical researchers.

Quickstart · Capabilities · Data sources · Deployment

Capabilities

• Literature question-answering - scan hundreds of papers and trial records at once to perform deep literature searches

• Data analysis via code execution - generate and execute code in a safely contained LangSmith Sandbox to perform almost any data analysis

• File and figure generation - create CSV and Excel files of data, Word docs such as clinical or lab protocols, and data visualizations and plots

Data sources

• PubMed - over 29 million scientific abstracts

• PMC full texts - full text of over 8 million open-access papers

• ClinicalTrials.gov - records from over 600,000 trials

• Web search - agentic search over the entire open web

• File upload - attach a spreadsheet, a reference-manager export (.nbib/.ris/.bib), a PDF, a compound set (.sdf/.smi), FASTA/GenBank, or a figure. Tables get analysed; a bibliography becomes a corpus the agent hydrates from PubMed and reads across

Quickstart

1. Get and configure LangSmith

You need a LangSmith account. Setup will prompt you to add your LANGSMITH_API_KEY.

LangSmith Sandboxes must also be enabled from the Sandboxes tab in your LangSmith console. On a personal account on the free Developer tier you will need to add a credit card to use sandboxes, but you get free 5 LangSmith Compute Units (LCUs) per month, enough for ~650 agent runs.

2. Get the code

In the terminal:

git clone https://github.com/langchain-samples/deep-life-sci.git

cd deep-life-sci

Needs git, which setup also uses to fetch the chat UI.

3. Get uv

uv is a widely-used package manager for Python that allows the setup script to install the necessary libraries.

On macOS / Linux:

curl -LsSf https://astral.sh/uv/install.sh | sh

On Windows:

irm https://astral.sh/uv/install.ps1 | iex

4. Run the setup script

uv run scripts/setup.py

The chat UI needs Node.js 20.9 or newer. If yours is missing or older, setup installs a private copy inside the repo; nothing else on your machine changes.

5. Configure your models

Setup will ask you how the agent should reach its models. Each answer has its own models file:

• LangSmith LLM Gateway (models.gateway.yaml). Model calls go through the LangSmith LLM gateway, so your workspace also needs the provider key behind them, added once in LangSmith under Settings → Integrations → Provider Secrets as OPENAI_API_KEY and/or ANTHROPIC_API_KEY. Add whichever providers the models you run use.
• Your own Anthropic or OpenAI API key (models.anthropic.yaml or models.openai.yaml). Model calls go straight to that provider, with the key you add during setup.
• Your own Amazon Bedrock credentials (models.bedrock.yaml). Setup asks for a Bedrock API key and a region. The models in that file run in us-east-1, us-east-2 and us-west-2; for other regions, see the Bedrock notes in it.

With your own credentials, LangSmith still handles tracing and sandboxes, so you still need its API key.

Setup records your answer as MODELS_FILE in .env. To change it later, set MODELS_FILE to another of these files and run setup again; it asks for any key that file needs.

Then, in that file, configure the models you want to use for the agent.

6. Run the agent

uv run scripts/dev.py                        # opens the chat UI in your browser (recommended)

# or

uv run agent "which papers base-edit PCSK9?" # runs headlessly in CLI

Ctrl-C to stop the running server.

Deploy to LangSmith

To run the agent in the cloud, with optional sign-in through your organization's identity provider, see DEPLOYMENT.md.

Coming soon

• Additional scientific data sources

Disclaimer

This is a demonstration project, intended for research and educational use. Its answers are generated by language models from published literature and trial registries, and may be incomplete, outdated, or wrong. It is not medical advice, and must not be used for clinical decision-making, diagnosis, or treatment.

Notes

Models: Deep Life Sci runs on GPT-5.6 Terra with High effort by default. To change the model or effort for the main agent, subagents or web search, edit your models file (models.gateway.yaml unless MODELS_FILE names another); changes apply to your next message. To add a provider, including a custom OpenAI-compatible endpoint, configure it in LangSmith under LLM Gateway. The wrench under the chat box shows what each role is running.

The chat UI in frontend/ began as agent-chat-ui.

Full text journal articles are only available if present in PMC's open-access subset. Only abstracts are available for paywalled papers.

MIT Licensed