philpaz

philpaz/recusal

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

34 good first / help-wanted issues · Python · last activity Sep 18, 2026

4 stars 0 forks 4 watchers Python Apache License 2.0

agent-governance agent-security agentic-ai ai-agents ai-governance ai-safety audit-log ci claude claude-code devsecops github-actions guardrails hooks llm llm-security python

View on GitHub Website

34 Open Issues Need Help Last updated: Sep 18, 2026

Open Issues Need Help

View All on GitHub

Wanted: gate LlamaIndex tool calls with Recusal 14 days ago

enhancement help wanted

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

Wanted: gate MCP tools in the OpenAI Agents SDK with Recusal 14 days ago

enhancement help wanted

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

Wanted: gate Pydantic AI tool calls with Recusal 14 days ago

enhancement help wanted

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

Wanted: gate CrewAI tool calls with Recusal 14 days ago

enhancement help wanted

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

Wanted: gate Google ADK tool calls with Recusal 14 days ago

enhancement help wanted

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

OpenAI Responses and Chat Completions example: gate each function call before your app runs it 17 days ago

enhancement help wanted

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

LangGraph pause, resume and replay: prove every resumed or replayed call is judged again 17 days ago

enhancement help wanted

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

agent_loop.py: a WHERE inside a SQL comment lets an unscoped DELETE through 17 days ago

bug good first issue

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

OpenAI Agents SDK example: gate function tools with a tool input guardrail, and refuse every tool it can't see 17 days ago

enhancement help wanted

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

LangGraph audit trail: record every gated tool call on a tamper-evident log, labeled langgraph.toolnode 17 days ago

enhancement help wanted

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

LangChain create_agent example: gate an agent's tool calls with the same recusal_gate, as middleware 17 days ago

enhancement good first issue

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

Every code block in the cookbook runs in CI, so a recipe can't silently drift 19 days ago

enhancement help wanted

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

Async LangGraph gate: the same check for graphs run with ainvoke, through awrap_tool_call 19 days ago

enhancement good first issue

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

Windows: concurrent first appends to a fresh audit log can fail with PermissionError instead of waiting 22 days ago

bug help wanted

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

LangGraph adapter example: gate a LangGraph agent's tool calls with Recusal 29 days ago

enhancement help wanted

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

recusal.checks: a NaN or out-of-range threshold silently passes the check 30 days ago

enhancement help wanted

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

Cookbook recipe 10 (compose policies) as a tested example, including fail-closed 30 days ago

documentation good first issue

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

Cookbook recipe 4 (wrong-subject guard): an unset active subject lets writes through 30 days ago

enhancement good first issue

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

Cookbook recipe 8 (quality gate) as a tested module: NaN coverage passes the gate 30 days ago

enhancement good first issue

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

SQL scope recipe: a DELETE or UPDATE inside a WITH clause is not checked about 1 month ago

enhancement good first issue

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

Cookbook recipe 1 (destructive shell) as a tested module: it misses what Recusal's own deny list refuses about 1 month ago

enhancement good first issue

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

Cookbook recipe 9 (approved-tool allowlist): a missing tool name is allowed through about 1 month ago

enhancement good first issue

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

Cookbook recipe 6 (injection quarantine) as a tested module: spacing and case slip past the phrase match about 1 month ago

enhancement good first issue

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

SQL scope recipe: a MySQL # comment can hide an unscoped DELETE about 1 month ago

enhancement good first issue

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

Workspace recipe: a project folder named like a secret refuses every write about 1 month ago

enhancement good first issue

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

Cookbook recipe 2 (WHERE on destructive SQL) as a tested module: comments and column names fool the keyword match about 1 month ago

enhancement good first issue

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

Cookbook recipe 3 (protect secrets, confine writes) as a tested module: a symlink escapes the workspace check about 1 month ago

enhancement good first issue

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

Egress recipe: the cookbook text lets a mixed recipient list through, and both versions trust a backslash URL about 1 month ago

enhancement good first issue

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

Add a `date_range` check to `recusal.checks` about 1 month ago

enhancement good first issue

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

Validate the enterprise managed-settings deployment patterns about 1 month ago

help wanted needs-environment

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

Validate the hook-timeout authorization outcome on a real deployment about 1 month ago

help wanted needs-environment

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

Documentation gap: no worked example of a custom `AuditSink` about 1 month ago

documentation good first issue

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

Cookbook recipe 7 (action budget) as a tested module about 1 month ago

enhancement good first issue

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python

Cookbook recipe 5 (egress allowlist) as a tested module about 1 month ago

enhancement good first issue

philpaz/recusal

4

Deterministic governance for Claude and MCP tool calls. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.

Python

#agent-governance#agent-security#agentic-ai#ai-agents#ai-governance#ai-safety#audit-log#ci#claude#claude-code#devsecops#github-actions#guardrails#hooks#llm#llm-security#python