replicant-ai 0.1.0

One command from research paper to running code

replicant

One command from paper to running code.

Give replicant an arXiv ID, a PDF, or a GitHub URL and it clones the repo, resolves the dependencies with AI, and drops you into a working Docker environment — without touching your local Python.

pip install replicant
replicant setup 2103.00020   # CLIP — OpenAI's vision-language model
replicant shell              # enter the environment

Works on any ML paper. GPU papers spin up an EC2 instance automatically.

Prerequisites

• Docker — must be installed and running
• AWS account — for Bedrock (the AI backbone) and optionally EC2 (cloud/GPU builds)

On first run, replicant setup will walk you through AWS credentials and model selection. You can also run it explicitly:

replicant init

The wizard checks Docker, installs Terraform if needed, verifies AWS credentials, and tests Bedrock access. Takes about 2 minutes.

Install

pip install replicant

Requires Python 3.9+.

Quick Start

# From an arXiv ID
replicant setup 2301.07041

# From a PDF
replicant setup ./attention-is-all-you-need.pdf

# From a GitHub URL directly
replicant setup https://github.com/karpathy/nanoGPT

# Paper doesn't include a GitHub link? Specify it manually
replicant setup 2301.07041 --github https://github.com/author/repo

# Enter the environment
replicant shell

# Enter a specific environment by ID
replicant shell a3f2c1b0

Commands

Command	Description
replicant init	Run the first-time setup wizard
replicant init --reset	Wipe config and re-run wizard
replicant setup <source>	Set up from arXiv ID, PDF path, or GitHub URL
replicant setup <source> --cloud	Build on AWS EC2 (GPU / large data)
replicant shell [env_id]	Enter environment (latest if no ID given)
replicant list	List all environments
replicant info [env_id]	Show environment details
replicant delete <env_id>	Remove environment and Docker image
replicant delete --all	Remove all environments
replicant validate [env_id]	Run post-build validation checks
replicant llm-config	Show current Bedrock config and test connection
replicant cloud status	List running cloud environments
replicant cloud teardown <env_id>	Shut down EC2 instance
replicant benchmark <corpus>	Batch-run across a CSV corpus of papers

Global flag: --verbose — stream build logs and debug output.

Cloud Execution

When a paper requires a GPU or downloads large datasets, replicant will prompt automatically:

GPU required. Run in the cloud? [y/N]:

Or pass --cloud to skip the prompt:

replicant setup 2103.00020 --cloud

This provisions a g4dn.xlarge EC2 instance (NVIDIA T4), builds the Docker image on it, pushes to ECR, and streams the shell over SSH. The instance runs until you explicitly tear it down:

replicant cloud teardown <env_id>

Cloud builds require Terraform — the replicant init wizard installs it automatically on macOS and Linux.

How It Works

replicant analyzes the repo and paper to build an environment spec, then generates a Dockerfile:

1. Detect — searches the repo for Dockerfile, environment.yml, requirements.txt, setup.py, pyproject.toml, Pipfile (3 levels deep, monorepo-aware)
2. Resolve — Claude reads the paper and repo together to fill in missing dependencies, fix version conflicts, and pin everything correctly
3. Validate — checks every package name against PyPI before building; phantoms are re-resolved automatically
4. Build — Docker image is built locally or on EC2; one-shot retry on failure re-resolves with the build error as context
5. Shell — drops you into /workspace with the cloned code mounted

All data lives under ~/.replicant/. Set REPLICANT_HOME to override.

Docs

• How it works — pipeline, AI dependency resolution, retry logic
• Supported environments — env file types, priority order, Dockerfile templates
• Troubleshooting — common failures and fixes

License

GNU Affero General Public License v3.0