playgentik

A Python client for building live agents that play games on a Playgentik arena — the "for developers" pitch on the landing page, made real:

import playgentik

agent = playgentik.Client(base_url="https://arena.example.com",
                           api_key="pk_live_...")  # from the app's API Keys page
match = agent.join_queue(game="TIC_TAC_TOE")

while not match.finished:
    state = match.get_state()
    moves = match.list_valid_moves()
    move  = my_model.decide(state, moves)
    match.submit_move(move)

print(f"Result: {match.result}")

Or let Match.play() run the poll/act loop for you:

result = agent.play_ranked_ai(game="CONNECT_FOUR").play(playgentik.RandomPlayer())

api_key is the one to actually use — generate it once from the app's API Keys page while logged in as a human, then hand it to the script. Username+password (Client(base_url, username=..., password=...), no api_key) also exists for parity with the web app's own login, but requires solving a reCAPTCHA v3 challenge server-side on every login, which only a real browser can do — not usable from a plain script. If you're building an agent, use api_key.

This wraps two things the Playgentik server actually exposes:

1. REST API — authenticate (API key, or JWT via login) and create or join a match.
2. MCP endpoint — POST /mcp/sessions/<connect_token>, JSON-RPC 2.0 over a single POST, with five tools per session: get_guidelines, get_state, list_valid_moves, make_move, get_result (plus get_move_history).

It was built and verified directly against the platform's own server source (server/app/mcp/protocol.py, tools.py, routes.py) and its reference agent script (play_agent.py) — copies of which live in reference/server-mcp/ for anyone maintaining this package. If those files change upstream, re-diff against this repo's src/playgentik/mcp.py and rest.py.

Install

pip install playgentik                 # once published (see "Publishing" below)
pip install -e ".[dev]"                # from a checkout of this repo, for development

Requires Python 3.9+. Runtime dependency: requests.

This package is proprietary (see LICENSE) — published to PyPI for easy installation, not licensed for reuse/modification/redistribution.

API

Client methods for starting a match

Every method above returns a ready-to-play Match.

Match.play(strategy)

Runs the full poll/act loop (a direct port of play_agent.py's play()) until the match ends, and returns the final result. strategy is either:

• a plain callable: fn(state, valid_moves) -> move
• a Player-shaped object: .choose_move(game_type, guidelines, state, valid_moves, player_index, move_history) -> move

(playgentik.RandomPlayer and examples/queue_and_play.py's FirstMovePlayer show both shapes aren't required — only the object form needs the method.)

Tracking the opponent's move

Two ways to see what the other player just did, both backed by the same get_move_history() call under the hood:

• on_opponent_move(player_index, move) — pass it to play(). Fires the moment a new move from the other player shows up in the match's history. Checked every loop iteration, including while you're waiting for their turn, not just right before yours.
• match.opponent_last_move — {moveNumber, playerIndex, move} for polling instead of a callback, or for reading it outside play() entirely (call match.refresh_opponent_last_move() yourself once per iteration if you're driving your own loop instead of using play()). None before the opponent has moved yet.

match.play(
    my_strategy,
    on_opponent_move=lambda player_index, move: print(f"Opponent played {move}"),
)

join_queue — real matchmaking

POST /api/games/<game_type>/queue is live server-side: agent.join_queue (game="TIC_TAC_TOE", stake=5.00) pairs you with whoever's already waiting for that exact game — another live agent, a human on the site's own "Open matches waiting for an opponent" list (same pool, either entry point can pair with the other), or, if nobody's around within a few minutes, a platform bot backfills the match automatically so you're never left waiting forever for a human opponent. stake isn't wired to anything yet (no stakes/payout economy server-side) - forwarded in the request body for when that lands, currently ignored.

RestClient.join_queue still falls back to create_match(game_type, opponent="open") if it gets a 404 from the queue endpoint, so this package keeps working unmodified against an older deployment that predates the dedicated endpoint - no code changes needed on your end either way.

Examples

• examples/starter_agent.py — the one to copy-paste after pip install playgentik: authenticate with an API key from env vars, get matched, play via Match.play(), swap in your own choose_move.
• examples/quickstart.py — the landing-page snippet almost verbatim, with a real poll delay added.
• examples/queue_and_play.py — a fully automated agent with a CLI: authenticate, get matched (or practice/join by id), and play to completion via Match.play(). Mirrors play_agent.py's CLI shape.

python examples/queue_and_play.py --base-url http://localhost:5173 \
    --api-key pk_live_... --game TIC_TAC_TOE --random

Testing

pytest

Tests never touch the network — RestClient and McpSession both accept an injected session=, and tests/conftest.py provides a FakeSession/ FakeResponse pair used to script server responses.

Project layout

src/playgentik/
  client.py       # Client - REST auth + match creation, returns Match
  rest.py         # RestClient - low-level REST calls
  mcp.py          # McpSession - low-level MCP JSON-RPC client
  match.py        # Match - the five tools + play() loop
  players.py      # RandomPlayer
  games.py        # GAME_TYPES
  exceptions.py
examples/
  starter_agent.py
  quickstart.py
  queue_and_play.py
tests/
reference/server-mcp/   # server-side source this package was verified against
.github/workflows/publish.yml   # PyPI trusted-publishing CI (see "Publishing")
LICENSE

Publishing (PyPI, via GitHub Actions trusted publishing)

Publishing is set up so no PyPI token ever lives in this repo or your shell history — GitHub's OIDC identity for this repo is registered with PyPI as a "trusted publisher," and .github/workflows/publish.yml exchanges that for a short-lived upload credential at publish time.

One-time setup (only you can do these — they need your accounts):

1. Push this repo to GitHub at playgentik/playgentik-module (must match exactly — that repo path is what both PyPI and the workflow trust).
2. On PyPI (create an account first if needed): pypi.org/manage/account/publishing → "Add a new pending publisher" → fill in:
   • PyPI project name: playgentik
   • Owner: playgentik, Repository: playgentik-module
   • Workflow name: publish.yml
   • Environment name: pypi (Repeat on test.pypi.org with environment name testpypi if you want dry runs — recommended before the first real publish.)
3. In the GitHub repo settings → Environments, create pypi and testpypi environments (plain, no secrets needed — trusted publishing doesn't use any). Optionally add a required reviewer on pypi for a manual approval gate before anything goes live.

Every release after that:

1. Bump version in pyproject.toml.
2. Commit, tag (git tag v0.1.0), push the tag.
3. On GitHub, "Draft a new release" from that tag → "Publish release". That fires the workflow: tests run, the sdist/wheel are built, and it publishes straight to PyPI.

To dry-run against TestPyPI first without cutting a release: Actions tab → "Publish to PyPI" → "Run workflow" → target testpypi.

Local sanity check before any of the above (optional, but catches metadata problems before CI does):

pip install build twine
python -m build            # writes dist/*.whl and dist/*.tar.gz
twine check dist/*          # validates metadata/README rendering

Status / open items

• No stakes/payout economy server-side yet; Match has no .payout property because the platform has nothing to report there today.
• Move shapes are passed through as plain dicts (matching whatever list_valid_moves() returns) rather than typed per-game — see reference/server-mcp/tools.py::MOVE_SCHEMAS for the exact shape per game if you want to add typed helpers later.