MLE-Agent
A CLI-oriented ML engineering assistant for building baselines, researching methods, and advancing local project workflows.
- Source repo
- MLSysOps/MLE-agent
- Stars
- ★ 1.6k
- Last updated
- 4mo ago
- License
- MIT
- Primary language
- Python
- FA score
- 39/100 · Major gaps
At a glance
- How it runs
- Works with
- Portable with changesOpenAI API · Claude API (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- An ML engineer who has a vague request, such as predicting stock prices from historical data, and wants to prototype and test a baseline locally.
- Not a fit if
- Users who want a GUI instead of a terminal-driven workflow
- Teams needing Hugging Face, S3, MLflow or W&B integrations out of the box
- Maintenance
- Commits in the last 6 monthslast commit 2026-07-10latest release 0.4.2 (2024-10-12)
- Source review
- 39/100 · Major gaps 2 safety controls not found
What does this agent do, and when should you use it?
MLE-Agent is presented as a pairing LLM agent for machine learning engineers and researchers. Its mle command-line interface creates and starts project directories and includes interactive terminal chat. The repository describes baseline construction, code execution and debugging on a local machine or cloud environment, plus research retrieval through Arxiv, Papers with Code, and general keyword search. It also documents Kaggle workflows, weekly-report generation, and local RAG support, while listing OpenAI GPTs, Anthropic Claude 3.5 Sonnet, and Ollama Llama3 among supported models. It fits ML teams that want one project-local workflow for planning, research, execution, and reporting, but the supplied material does not document provider credentials or model configuration.
After mle new <project name> creates a directory, mle start begins the ML/AI project workflow and mle chat opens an interactive terminal session. The project describes building ML/AI baselines from requirements, retrieving methods through Arxiv, Papers with Code, and general keyword search, and executing, debugging, and fixing code on a local machine or cloud environment. mle kaggle starts a Kaggle task, while mle kaggle --auto accepts dataset paths, a description, submission file, submission example, and competition ID for an automated workflow spanning data preparation, training, and debugging. For reporting, mle report starts a local web application at http://localhost:3000/ for GitHub-based reports, and mle report-local --email=<git email> --start-date=YYYY-MM-DD --end-date=YYYY-MM-DD <path_to_git_repo> creates a report from a local Git repository.
- An ML engineer who has a vague request, such as predicting stock prices from historical data, and wants to prototype and test a baseline locally.
- A Kaggle participant who has joined a competition and has the datasets and submission artifacts ready for the parameterized
mle kaggle --autoworkflow. - A researcher planning an ML approach who wants to consult Arxiv, Papers with Code, and keyword-search results alongside implementation work.
- A developer maintaining an existing ML project who wants a terminal-based
mle chatinteraction mode within that project directory. - A team member who needs a weekly work report covering development progress, communication notes, references, and to-do items from GitHub or a local Git repository.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
pip install -U mle-agentuv pip install -U mle-agentpip install -e
How do you install or deploy this agent?
Install from PyPI:
pip install -U mle-agentOr with uv:
uv pip install -U mle-agentInstall from source:
git clone https://github.com/MLSysOps/MLE-agent.git
cd MLE-agent
uv venv .venv
source .venv/bin/activate
pip install -e .The supplied material lists support for providers including OpenAI, Anthropic, and Ollama, but does not document required credentials, environment variables, or model-configuration steps.
How do you use this agent?
Create and start a project:
mle new <project name>
cd <project name>
mle startStart terminal chat in that project:
mle chatStart a Kaggle workflow:
mle kaggleGenerate a report from a local Git repository:
mle report-local --email=<git email> --start-date=YYYY-MM-DD --end-date=YYYY-MM-DD <path_to_git_repo>What are this agent's strengths and limitations?
- Combines ML baseline creation, code execution and debugging, research retrieval, and project guidance in one project-level CLI workflow.
- Explicitly integrates Arxiv, Papers with Code, and general keyword search for bringing research references into implementation planning.
- Provides a parameterized
mle kaggle --autoflow that accepts practical competition inputs such as datasets, descriptions, and submission artifacts. - Includes terminal chat, a GitHub-report web entry point, and a local-Git reporting command rather than focusing only on coding.
- The README lists OpenAI GPTs, Anthropic Claude 3.5 Sonnet, and Ollama Llama3, indicating a multi-model integration direction.
- The supplied material does not document provider credentials, environment variables, configuration files, or a default model, so first-run setup requires further verification.
- The Kaggle automation mode requires that the user has joined the competition and prepared datasets, submission files, and a submission example.
- Its documented local filesystem and code-execution capabilities require evaluating the execution boundary for project code and data before adoption.
- Several integrations, including cloud data and testing platforms, Hugging Face, AWS S3, MLflow, and W&B, remain unchecked roadmap items.
- Although the README describes end-to-end Kaggle task completion, the supplied evidence provides no benchmark, success-rate, or resource-consumption data.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| MLE-Agent This agent | 39 · Major gaps | CLIFree + model costs | ★ 1.6k | 4mo ago | Python | — |
| AgentSwarms Data and Agent Platform | 79 · Good | Web appFree + model costs | ★ 302 | today | TypeScript | OpenAI API |
| optim-agent | 68 · Some gaps | Library / SDKFree + model costs | ★ 796 | 3mo ago | Python | Codex · Claude Code |
| Gyoshu Research Lab | 44 · Major gaps | Agent plugin / skillFree + model costs | ★ 243 | 9mo ago | TypeScript | Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The CLI, project files, reporting, and data workflows can access local files; the report API also accepts an optional token. The supplied evidence does not show permission isolation, secret protection, or a complete data-flow account, so least privilege, sensitive-data handling, and transparency are only partial. Interactive modes and the Kaggle membership requirement provide limited human control, but auto mode can complete tasks independently and no side-effect confirmation or recovery path is shown. Arxiv and Papers with Code are named as sources, while attribution for other generated content is unclear.
The project specifies dependencies and includes some model tests, supporting partial judgments of consistency and availability. However, README claims are not individually corroborated by the supplied code, dependency ranges are broad, and comprehensive fallback behavior is not shown. Some CLI paths report exception type, location, and message; other paths propagate exceptions, so failure handling is partial.
The README identifies ML engineers and researchers and describes baseline, reporting, Kaggle, and chat scenarios. It claims end-to-end autonomy without clearly defining where autonomy stops, which actions need review, or other capability boundaries. Installation guidance covers PyPI, source, and a Linux/macOS virtual environment, but leaves some environment requirements unclear.
The README is organized around installation, usage, scenarios, roadmap, contribution, and licensing, with command examples and dated milestones. Installation instructions contain a duplicated pip alternative and limited Python/platform detail. pyproject.toml declares Apache-2.0 while LICENSE and README state MIT, creating a material license conflict. Maintainers, community, and contribution routes are visible, but no clear maintenance responsibility or update commitment is stated.
Command examples and reporting/Kaggle workflow descriptions make intended outputs actionable. Code retrieval, research search, and automated debugging could save time for the target audience. The README makes broad capability claims while the dependency set and configuration create setup costs; the supplied sources do not quantify benefits, costs, or required human review.
Version, dependencies, commands, and some tests make a subset of implementation claims traceable, but many README feature claims are not corroborated by the supplied source files. README, package metadata, and license file conflict on MIT versus Apache-2.0. The documentation does not consistently separate implemented capabilities, roadmap items, and unverified outcomes, so verifiability is partial.
- Not found in source: disclosed external effectsEstablish which external systems it writes to, sends to or changes, and verify with test accounts or repos before production.
- Not found in source: rollback or recovery pathBack up first, or work on a git branch or snapshot, so its changes can be undone.
- Package metadata declares Apache-2.0, while the README and LICENSE state MIT; verify the applicable license before use or redistribution.
- The tool processes local project files and connects to model or external services; the supplied materials do not fully explain data sent or secret protection.
- The README claims broad automation capabilities, but the supplied code is only a sample and does not establish every workflow or outcome.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
What's Changed * [MRG] build version v0.4.0 by @huangyz0918 in https://github.com/MLSysOps/MLE-agent/pull/208 * [MRG] update readme for 0.4.0 by @HuaizhengZhang in https://github.com/MLSysOps/MLE-agent/pull/209 * [MRG] deepseek support by @umnooob in https://github.com/MLSysOps/MLE-agent/pull/212 * [MRG] add kaggle requirement summary by @leeeizhang in https://github.com/MLSysOps/MLE-agent/pull/213 * [MRG] add langfuse by @leeeizhang in https://github.com/MLSysOps/MLE-agent/pull/214 …View on GitHub ↗ -
Release notes excerpt
What's Changed * [MRG] 👷 build: 0.3.1 by @huangyz0918 in https://github.com/MLSysOps/MLE-agent/pull/124 * [MRG] add star history by @HuaizhengZhang in https://github.com/MLSysOps/MLE-agent/pull/125 * [MRG] Improve advisor agent with markdown output by @huangyz0918 in https://github.com/MLSysOps/MLE-agent/pull/132 * [MRG] Github Integration - Information collector by @huangyz0918 in https://github.com/MLSysOps/MLE-agent/pull/134 * [MRG] Upgrade setuptools by @huangyz0918 in https://github.com/MLSysOps/MLE-agent/pull/135 …View on GitHub ↗ -
Release notes excerpt
What's Changed * [MRG] build version v0.3.0 by @huangyz0918 in https://github.com/MLSysOps/MLE-agent/pull/120 * [MRG] add previous step&task resume by @leeeizhang in https://github.com/MLSysOps/MLE-agent/pull/123 Full Changelog: https://github.com/MLSysOps/MLE-agent/compare/0.3.0...0.3.1View on GitHub ↗
FAQ
Which model-provider credentials do I need?
Does it access or execute local content?
Can it run a Kaggle competition fully automatically?
mle kaggle --auto, but requires that you have joined the competition and have datasets, a description, submission file, submission example, and competition ID ready.What does the reporting feature produce?
Is MLE-Agent free?
Is MLE-Agent open source?
Is MLE-Agent safe to use?
Is MLE-Agent still maintained?
Who makes MLE-Agent?
Thanks — we use this to improve the page.