Skip to content
Hosted service in public preview. 3.0 installs from GitHub until it is on PyPI.

Releases

This page lists the releases and version history for agenticml-lite.

Use this table to choose the version for your setup:

Path Version Install command
Local tracking only 1.0.0 from PyPI pip install agenticml-lite
Local tracking only 3.0 from GitHub pip install 'agenticml-lite[system] @ git+https://github.com/agenticML/agenticml-lite@dev'
Self-hosted server 3.0 from GitHub pip install 'agenticml-lite[system] @ git+https://github.com/agenticML/agenticml-lite@dev'
Hosted service 3.0 client from GitHub pip install 'agenticml-lite[system] @ git+https://github.com/agenticML/agenticml-lite@dev'

pip install agenticml-lite installs 1.0.0 from PyPI. It supports local tracking only. It cannot sync to a 2.x or 3.0 server.

Version 3.0 includes the server and hosted service support. It is a pre-release and is not on PyPI yet. Install it from GitHub. This command follows the dev branch, so what you get moves with development.

A 2.2 client has not been checked against the hosted service. Use the 3.0 client.

To check which version you have installed:

Terminal window
pip show agenticml-lite

This works on every version. The agenticml --version command works from 3.0. On 1.0.0 and 2.x, use pip show.


Status: Pre-release. The hosted service is in public preview.

Version 3.0 is on the dev branch. It is not tagged and is not on PyPI yet. The pyproject.toml version there is 3.0.0.

Install from GitHub:

Terminal window
pip install 'agenticml-lite[system] @ git+https://github.com/agenticML/agenticml-lite@dev'
  • Database storage: Server state lives in a SQL database. SQLite is used for single-process self-hosted deployments. PostgreSQL is used for multi-process or hosted deployments.
  • Workspaces: All runs, API keys, and memberships belong to isolated workspaces on a single hostname. People have a role in each workspace. Dashboard views run at /w/<slug>/.
  • Workspace switcher: Switch between workspaces from the dashboard.
  • Hosted service: A service on AWS at https://app.agenticml.xyz. It is a public preview, run on a best-effort basis with no SLA. See the hosted quickstart.
  • Firebase sign-in: Sign in to the hosted service with Google or email and password.
  • Self-serve sign-up: See Sign-up for whether open sign-up is on.
  • Free plan limits (as of Nov 13, 2026): 5 GB storage, 2 GB uploads per UTC day, 600 requests per minute per workspace, 5 people (members plus pending invites), and 1 free workspace per account. Over the storage, daily upload or people limit the server answers 403. Over the request rate it answers 429 with Retry-After.
  • Several server processes and S3 objects: PostgreSQL deployments can run more than one server process. Content-addressed objects (code snapshots, media, copied artifacts) can be stored in Amazon S3. The hosted service uses both.
  • Metric summaries: The server keeps a summary of each metric series in the database, so the runs list can sort and filter by metric.
  • Automatic data conversion: A 2.x data directory converts to the 3.0 database on first start. Converted runs belong to the default workspace.
  • Client support window: By default the server requires client version 2.2 or newer. Older clients get HTTP 426 and must update.

Status: Shipped in the 2.2 code on main; never published on its own; included in 3.0.

Version 2.2 (commit bd0d499 on dev) reached main on Nov 7, 2026, as merge e0916d7. It was never tagged and never published on PyPI.

  • Users and roles: Added admin, member, and viewer roles. Admins manage users and keys. Members sync runs and edit notes. Viewers have read-only access.
  • Per-user keys: API keys belong to individual users. Runs record user ownership in remote.json.
  • Dashboard sign-in: Sign in with password credentials and one-time invite links. Supports single sign-on through a trusted identity proxy (x-amzn-oidc-identity, Tailscale-User-Login, or X-Forwarded-Email).
  • Audit log: Append-only log at <data>/server/audit.jsonl records logins, user changes, key changes, and note edits.
  • Team administration: Dashboard page for admins to manage users, roles, and keys.

If you run 2.2 from Git, the 3.0 server pages do not describe your server. The 2.2 documentation stays in the repository. See the 2.2 documentation tree on GitHub.


Status: Shipped in the 2.2 code on main; never published on its own; included in 3.0.

Version 2.1 was never tagged and never published on PyPI.

  • Lineage replication: The client uploads Git bundles of shadow commits. This enables code diffs between runs in the server dashboard.
  • By-reference artifacts: Artifacts logged with copy=False can be uploaded by reference using agenticml sync --artifacts.
  • Operator commands: agenticml server status inspects runs, storage, and keys. agenticml server verify validates file sizes and re-hashes objects.
  • Server up command: agenticml server up provisions an internet-accessible server through Tailscale Funnel.

Status: Shipped in the 2.2 code on main; never published on its own; included in 3.0.

Version 2.0 was merged to dev in pull requests from Nov 3 to Nov 5, 2026, and reached main with 2.2 on Nov 7, 2026. It was never tagged and never published on PyPI.

  • Self-hosted server: agenticml server serve --data DIR runs a central sync server and web dashboard.
  • Sync protocol v2: Byte-for-byte run replication over HTTP. Interrupted syncs resume where they stopped.
  • Background sync: agenticml sync --follow keeps syncing active and new runs in the background.
  • Server dashboard: Web interface displays runs, charts, lineage, and run sync status. Includes key-based login and the Keys management page.
  • Systemd units: Example unit files for running the server and background sync client as system services.
  • Client commands: agenticml login --host URL, agenticml whoami, and agenticml logout work against the new server.
  • Old sync removed: Sync to the old hosted service (api.agenticml.xyz) was removed.
  • No default host: agenticml login needs --host the first time.
  • Python 3.9 dropped: Requires Python 3.10 or newer.
  • No Docker deployment: The server ships only as a Python package. Docker images and compose files were removed.
  • POSIX only: The server requires POSIX file locking (fcntl). It runs on Linux and macOS. Windows runs the client and local agenticml ui.

Status: Released Nov 1, 2026 (tag v1.0.0). On PyPI.

Version 1.0.0 is the only release published on PyPI.

  • Local-first tracking: Log machine learning experiments directly to disk in ~/.agenticml. No server or database required.
  • Flexible logging: Track runs from Python with agenticml.init(), agenticml.log(), and agenticml.finish(), or wrap training scripts without code changes using agenticml run train.py.
  • Automatic capture: Captures command-line arguments, Git code state and PyTorch components. agenticml run also mirrors TensorBoard summaries.
  • Local dashboard: agenticml ui serves a web dashboard at http://127.0.0.1:6006 with run tables, metric charts, lineage graphs, and code diffs.
  • CLI tools: Manage the local store with agenticml ls, agenticml lineage, agenticml du, and agenticml gc.
  • Distributed training: Multi-GPU support with rank-0 logging.
  • Old hosted sync retired: The login and sync commands in 1.0.0 default to https://api.agenticml.xyz, a service that has been retired. Sync from 1.0.0 no longer works, and old guides that use that host no longer apply. Version 1.0.0 cannot sync to a 2.x or 3.0 server.