maigp-agent-core

Shared agent-governance lifecycle for every MAIGP framework adapter.

Part of MAIGP — the Mediated AI Governance Protocol: a runtime consent / scope / evidence layer for AI systems. Every governed call is CHECKed before it runs (policy, scope, jurisdiction, circuit-breaker) and RECORDed + TRACEd as tamper-evident evidence after.

What it governs

Governs the AIGP agent loop itself — the primitive that all framework adapters build on.

Governance model

This adapter maps the framework's lifecycle onto the AIGP agent-governance loop provided by maigp-agent-core:

Install

pip install maigp-agent-core

Quick start

from aigp_agent_core import AgentGovernance

gov = AgentGovernance(gov_url="https://gov.example.com",
                      app_id="my-agent", hmac_secret="...")

# Start of an agent run
gov.pre_invoke(agent_name="analyst", model_id="claude-opus-4", user_id="u1")

# Agentic control points (RFC-agentic): gate tools, plans, escalations
decision = await gov.on_tool_call("search", params={"q": "..."})
await gov.plan_submit(steps=[{"tool": "search"}, {"tool": "write"}])
await gov.escalate(reason="budget exceeded", action={"type": "pause"})

# Token accounting + completion
gov.on_model_call(input_tokens=500, output_tokens=200)
await gov.step_complete(step_number=1, status="SUCCESS")
await gov.post_invoke(status="SUCCESS")

Who should use this

You normally install a framework adapter (e.g. maigp-crewai, maigp-microsoft-agent-framework) which depends on this package. Install maigp-agent-core directly only when writing a custom adapter for a framework not yet covered.

Control points

pre_invoke · on_model_call · on_tool_call · plan_submit · escalate · attest_input · step_complete · post_invoke · on_error.

Links

• Protocol, specs & full adapter catalog: https://github.com/owner-spec/aigp-protocol
• Issues: https://github.com/owner-spec/aigp-protocol/issues

License: Proprietary. © Kanjani AI Research & Causum.