Autoresearch Swarm
Run reproducible research loops through an API, CLI or coding agent. Propose a change, execute it in isolation, evaluate it, then keep or reject it using recorded evidence. Multiple agents and workers can explore candidates concurrently.
The default runtime uses Python, asyncio, SQLite and local artifact storage. Temporal, Cloudflare and remote workers extend the same contracts. A UI is optional.
Inspired by Andrej Karpathy's autoresearch. The platform implementation is original and MIT-licensed. Upstream source is not bundled or relicensed.
Start locally
From this checkout, with Python 3.11 or newer and uv:
uv sync --group dev
uv run research init
uv run research run examples/cpu_bigram
The CPU example trains a small model and evaluates it independently. See local runs for budgets, resumption and campaign reports.
Expose the same control surface to an agent:
uv run research capabilities
uv run research mcp
Agents can propose candidates, submit trials, inspect results and request promotion. The engine owns execution state, budgets and evaluator authority. Start with agent control, the Python SDK or the HTTP API.
Examples
| Example | Purpose |
|---|---|
| CPU bigram | A small training and evaluation loop. |
| Synthetic pose | Adopt an ordinary research repository and replay archived receipts. |
| MLX workload | Native Apple Metal training with the reference scorer. |
| Pinned upstream workload | Preserve the original fixed-budget experiment contract. |
| Policy learning | Train, evaluate and explicitly adopt a research policy. |
Examples distinguish integration checks from measured scientific improvement. Qualification status records the tested platforms, workloads and provider runtimes. Native CUDA parity and deployed Cloudflare operation require separate qualification.
Guides
| Task | Guide |
|---|---|
| Add a research repository | Project adoption, manifest, workload protocol |
| Control a research loop | Campaigns, evaluation and promotion |
| Register compute | Remote workers, engine profiles, worker control |
| Change infrastructure | Orchestration, Temporal, Cloudflare, R2 |
| Preserve and recover evidence | Artifacts and source, historical import, backup |
| Extend the platform | Adapter contracts, design contract, protocol conformance |
| Prepare a release | Release previews and PyPI publishing |
Development
uv run ruff check .
uv run ruff format --check .
uv run mypy src
uv run pytest
See CONTRIBUTING.md for setup, scoped testing and review.
1 maintainer
Metadata
1 maintainer
Release files for autoresearch-swarm 0.1.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| autoresearch_swarm-0.1.0.tar.gz | 578.1 kB | With provenance | Details for autoresearch_swarm-0.1.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| autoresearch_swarm-0.1.0-py3-none-any.whl | Python 3 | none | any | Details for autoresearch_swarm-0.1.0-py3-none-any.whl |
Total release size: 1.2 MB
Release files / autoresearch_swarm-0.1.0.tar.gz
| Download URL | autoresearch_swarm-0.1.0.tar.gz |
|---|---|
| Size | 578.1 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
71f282019f9002beba7b5a5e3ce5824f06744fb1af1971b7c8f8c5dd3b7a824e
|
|
BLAKE2b-256 checksum How to use checksums |
e32969171b1bb4962d50550b605a08a7bec2b27ea1418529b182369f94de1f17
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 1, 2026.
Transparency logRelease files / autoresearch_swarm-0.1.0-py3-none-any.whl
| Download URL | autoresearch_swarm-0.1.0-py3-none-any.whl |
|---|---|
| Size | 616.7 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
620fd02a83e12b80019fb0b9ba801b48f0274463179fe90d7e4a3b341c9392f1
|
|
BLAKE2b-256 checksum How to use checksums |
84964f65061cd5a50ef9172e9219348d91a7d976d1e6e3c16c4890f71fa0cd3f
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 1, 2026.
Transparency logSecurity for autoresearch-swarm 0.1.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
Full provenance
All 2 files in this release have been uploaded with provenance metadata.
Attestation for autoresearch_swarm-0.1.0.tar.gz
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 1, 2026.
Transparency logA note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.
