Each adapter package connects acgs-lite>=3.0.1,<4 rule evaluation to a single agent framework or MCP server. All 13 packages are published on PyPI as Alpha releases under Apache-2.0 and fail closed when governance evaluation raises an error.
Two integration mechanics
Not every framework hook owns the tool call.
Before choosing an adapter, check whether the target framework exposes a middleware wrapper around execution or only a pre-call veto callback.
Shared foundation: all 13 adapters evaluate rendered tool-call arguments through ToolCallGovernor on acgs-lite>=3.0.1,<4.
These five adapters sit directly around the tool dispatch function (such as FastMCP's call_next or LangChain's wrap_tool_call). Because the middleware controls when and whether the underlying callable is invoked, it can enforce pre-call checks, interrupt for human approval, and inspect post-call outcomes on the governed path.
These eight adapters plug into host callbacks (such as Claude Agent SDK's PreToolUse permission decision, OpenAI Agents input guardrails, or Google ADK's before_tool_callback). They can allow, deny, or request review before the tool runs, but the host framework owns the actual tool dispatch after an allow decision.
Complete package ledger
Choose the adapter for your agent stack.
Every entry below lists the exact published PyPI version, the inspected hook symbol from the repository source, the supported framework constraint, and the integration caveats to verify before deployment.
Execution boundary model
Showing 13 of 13 Alpha packages · All require acgs-lite>=3.0.1,<4 · Apache-2.0
Boundary note: Provider-hosted tools (such as hosted web search or file search) execute outside local Python hooks and are not governed by this adapter.
Boundary note: Pre-0.2 framework target; tested against beeai-framework 0.1.85 with a scripted model on Python >=3.11,<3.14.
Model Context Protocol
Why the MCP server boundary reaches every client.
When tools are exposed through a FastMCP server, placing governance at ACGSMiddleware.on_call_tool governs requests regardless of which agent framework or IDE client initiates the call.
Illustrative setup for fastmcp-acgs==0.1.0. Evaluates tool arguments against acgs-lite rules before call_next runs. Does not mint gove-zone kernel receipts.
Scope & verification
Know what the adapters cover today.
Each adapter repository is an independent Python package with its own unit test suite and CI matrix. Keep these four boundaries in view during evaluation.
All 13 adapters ship the ACGS-Lite ToolCallGovernor path: they serialize tool names and arguments, evaluate configured keyword/regex rules, and fail closed if governance evaluation raises an exception.
02
Receipts, grants, and EffectStore are separate layers
In the current 0.1.x adapter releases, 0 of 13 packages directly wire SignedReceipt, ExecutionGrant, or EffectStore. If you require receipt-gated tool execution or SQLite effect reservation, integrate those contracts explicitly in your tool executor.
03
Version pin coupling with ACGS-Lite 3.0.1
Every adapter declares acgs-lite>=3.0.1,<4 in pyproject.toml (and pins ==3.0.1 in oldest-version CI checks). Do not pair these adapters with older 2.x acgs-lite releases.
04
Bypass paths and provider-hosted tools
An adapter only governs tool calls that pass through its registered hook or middleware. Direct Python function calls, sub-agents without the hook attached, or provider-hosted server tools (for example in openai-agents-acgs) bypass local interception.
Put one real workflow under your control.
Choose one tool. Define what it may do. Inspect what happens when a request crosses the boundary.