Fly Labs
Ask FlyBot

A context-aware permission guard for Claude Code

We needed something like --dangerously-skip-permissions that doesn’t nuke your untracked files, exfiltrate your keys, or install malware.Claude Code's permission system is allow-or-deny per tool, but that doesn’t really scale. Deleting some files is fine sometimes. And git checkout is sometimes not fine. Even when you curate permissions, 200 IQ Opus can find a way around it. Maintaining a deny list is a fool's errand.nah is a PreToolUse hook that classifies every tool call by what it actually does, using a deterministic classifier that runs in milliseconds. It maps commands to action types like filesystem_read, package_run, db_write, git_history_rewrite, and applies policies: allow, context (depends on the target), ask, or block.Not everything can be classified, so you can optionally escalate ambiguous stuff to an LLM, but that’s not required. Anything unresolved you can approve, and configure the taxonomy so you don’t get asked again.It works out of the box with sane defaults, no config needed. But you can customize it fully if you want to.No dependencies, stdlib Python, MIT.pip install nah && nah installhttps://github.com/manuelschipper/nah

Hacker NewsToolAISource
0Sign in to voteCopy link

FL score

76

out of 100

Verdict

VALIDATE

high confidence

Competition

6

competitors found, emerging market, funded players

Trend

No signal yet

A context-aware, deterministic permission guard for AI coding agents that addresses a critical and growing security vulnerability.

The pain

AI coding agents lack reliable, fine-grained permission controls, making them susceptible to exfiltrating sensitive data, installing malware, or making dangerous system changes. Current solutions rely on unreliable LLM instructions or are not designed for the semantic layer of agent actions.

The gap

While many funded companies focus on general AI security, there's a specific gap for a deterministic, context-aware permission guard that acts as a pre-tool-use hook, semantically classifying actions (e.g., filesystem_read, git_history_rewrite) and applying explicit policies, unlike the ambiguous LLM-based approaches or broader security tools.

Build angle

Develop 'nah' as a highly-optimized, 'stdlib Python' pre-tool-use hook for Claude Code (and potentially other agents) that uses a deterministic classifier to identify specific action types and enforce granular, context-aware policies (allow, context, ask, block) by default, with optional LLM escalation for ambiguous cases.

Strengths

  • Addresses a critical and growing security pain point in the AI agent space.
  • Deterministic approach offers a clear advantage over unreliable LLM-based permissioning.
  • Specific, narrow angle targeting a clear user need.
  • Strong market signals for willingness to pay (significant funding in the space).
  • Good initial MVP demonstrated by existing GitHub project.

Risks

  • The 'deterministic classifier' is technically complex to build comprehensively and maintain for all possible tool calls and contexts.
  • Highly competitive landscape with several well-funded players tackling AI security.
  • Achieving widespread adoption might be challenging against established security vendors.
  • Long-term solo scalability could be an issue for maintaining the classifier against evolving threats and agent capabilities.

Next steps

  1. Actively seek feedback from current 'nah' users and Claude Code developers regarding classifier accuracy and coverage.
  2. Validate the most critical security vulnerabilities users want to prevent with specific examples.
  3. Explore partnerships or integrations with existing AI agent frameworks to broaden reach and validate the pre-tool-use hook model.
  4. Investigate monetization models beyond the current open-source offering, focusing on team/enterprise features.

Questions about this idea?

FlyBot reads the scoring and gives you a second opinion on “A context-aware permission guard for Claude Code”.

Open FlyBot