Ai2 Soft-Verified Efficient Repository Agents (SERA) Claude Code Proxy

This repo allows Claude Code to be used with the Ai2 Open Coding Agents SERA model.

You will need Claude Code and uv installed to set up the SERA CLI.

For more information about Open Coding Agents and SERA, see:

• sera-cli Demo on YouTube
• SERA Data Generation and Training Code
• SERA Technical Report
• Ai2 Open Coding Agents Blog Post
• Ai2 Open Coding Agents Hugging Face Collection

Quick Start with Modal

The fastest way to try SERA is with Modal, which handles GPU provisioning, vLLM deployment, and downloading the model automatically. This takes ~10m for the first run as ~65GB of model weights are downloaded. Subsequent runs will cache the model and start up faster.

When you exit Claude Code, the Modal app will automatically get cleaned up.

# Install modal and sera globally
uv tool install modal
uv tool install ai2-sera-cli

# Setup modal (this will prompt you to set up an account)
modal setup

# Deploy SERA to Modal and launch Claude Code this uses allenai/SERA-32B by default
sera --modal

# Use the allenai/SERA-8B model. Non-SERA models are untested and may not behave as expected
sera --modal --model allenai/SERA-8B

Using Existing Endpoints

If you have an existing vLLM endpoint for the SERA model (e.g., from a shared deployment or your own infrastructure):

# Install sera globally
uv tool install ai2-sera-cli

# Set the API key if your endpoint requires authentication
export SERA_API_KEY=<your API key>

# Run sera with your endpoint
sera --endpoint <endpoint URL>

Shared Deployments with deploy-sera

For teams or multi-user setups, you can create a persistent vLLM deployment on Modal using deploy-sera. Unlike sera --modal which creates ephemeral deployments that stop when you exit, deploy-sera creates persistent deployments that stay up until explicitly stopped.

# Deploy a persistent vLLM instance with your choice of model
deploy-sera --model allenai/SERA-32B
deploy-sera --model allenai/SERA-8B

# The command outputs an endpoint URL and API key
# Share these with your team members

# Team members can then connect with:
SERA_API_KEY=<api-key> sera --endpoint <endpoint-url>

# Stop the deployment when done
deploy-sera --stop

deploy-sera Options

Deploying Private Models

For private models (e.g., fine-tuned on a proprietary codebase), use --hf-secret to authenticate with HuggingFace:

# 1. Create a Modal secret with your HuggingFace token
modal secret create huggingface HF_TOKEN=hf_your_token_here

# 2. Deploy your private model
deploy-sera --model your-org/private-sera-model --hf-secret huggingface

# 3. Users connect with the provided endpoint and API key
SERA_API_KEY=<api-key> sera --endpoint <endpoint-url>

For ephemeral single-user deployments, the same --hf-secret flag works with sera --modal.

Self-Hosted vLLM

You can run SERA with vLLM on any cloud GPU provider or your own hardware directly with vLLM.

On the server:

python -m vllm.entrypoints.openai.api_server \
    --model allenai/SERA-32B \
    --host 0.0.0.0 \
    --port 8000 \
    --max-model-len 32768 \
    --tensor-parallel-size 2 \
    --trust-remote-code \
    --enable-auto-tool-choice \
    --tool-call-parser hermes

On your dev machine:

uv tool install ai2-sera-cli
sera --endpoint http://your-server:8000/v1/chat/completions

Configuration

sera CLI Options

Environment Variables

API Key Authentication

The proxy supports API key authentication for vLLM endpoints:

• sera --modal: API key is auto-generated and managed in the background
• deploy-sera: API key is auto-generated and printed so it can be shared with team members
• Existing endpoints: Set SERA_API_KEY environment variable before running sera
• Self-hosted vLLM: Start vLLM with --api-key YOUR_KEY, then set SERA_API_KEY=YOUR_KEY

The proxy includes the API key in the Authorization: Bearer <api_key> header when making requests.

Citation

@misc{shen2026serasoftverifiedefficientrepository,
      title={SERA: Soft-Verified Efficient Repository Agents},
      author={Ethan Shen and Danny Tormoen and Saurabh Shah and Ali Farhadi and Tim Dettmers},
      year={2026},
      eprint={2601.20789},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2601.20789},
}