marbots-sdk (Python)

Typed, zero-dependency Python client for Marbots, the multi-agent collaboration platform. Responses are frozen dataclasses; kernel packs, permission profiles, event types and model settings are enums/helpers, so mypy (and your IDE) catch typos before you run anything.

pip install marbots-sdk

from marbots_sdk import MarbotsClient, BotSpec, KernelPack, PermissionProfile, ModelRef, EventType

mb = MarbotsClient("http://localhost:5170")            # api_key="..." if the server requires one

# Every bot can use its own model; ModelRef.DEFAULT follows the workspace default.
sari = mb.bots.create(BotSpec(
    name="Sari", role="UX designer",
    kernel_functions=[KernelPack.FILES, KernelPack.WEB],
    permission_profile=PermissionProfile.WORKSPACE_WRITE,
    model=ModelRef.of("azure", "gpt-5.6-luna"),
))
mb.bots.set_model("atlas", ModelRef.DEFAULT)
print(mb.models.list().default)                         # e.g. azure/gpt-5-mini

thread = mb.threads.create(sari.id)
result = mb.threads.send(thread.id, "Sketch a landing-page wireframe", wait=True)
print(result.task.model, result.task.cost_usd, result.text)

for event in mb.events.stream(thread.id):               # live Server-Sent Events
    if event.type == EventType.TOOL_CALL_STARTED:
        print("tool:", event.message)
    if event.is_task_finished():
        break

Type safety is tested: python tests/typecheck/run.py requires mypy --strict to pass on the SDK and to reject every deliberate typo in tests/typecheck/check_types.py. The conformance tests (tests/test_conformance.py) run against a real Marbots server.

Built by Gravicode Studios, led by Kang Fadhil. MIT license.

Computers, placement and learning evaluation (0.2.0)

See docs/en/computers.md and the Skills guide.

from marbots_sdk import AUTO_HOST, BotSpec, ContainerProfile, KernelPack

# Computers (agent hosts): list, one-time enrollment token, SSH bootstrap (credentials used once)
for h in mb.agent_hosts.list():
    print(h.name, h.status, h.capabilities)
print(mb.agent_hosts.create_enrollment("design-pc").enroll_command)

# A bot that runs on whichever computer fits, with its shell in Docker and parallel sub-agents
nova = mb.bots.create(BotSpec(name="Nova", host_ref=AUTO_HOST, container=ContainerProfile(image="python:3.12-slim"),
                              kernel_functions=[KernelPack.FILES, KernelPack.SHELL, KernelPack.SUBAGENTS]))

# Learning evaluation
for e in mb.skills.evaluations():
    print(e.name, e.version, e.verdict, e.reason)
mb.skills.set_auto_rollback(True)