AI agents install packages. Attach checks them first.
attach-guard-demo.mp4
Claude Code installs packages on your behalf — often without you reviewing each one. Existing security tools scan after the fact or rely on advisory prompts that Claude can skip. There is no open-source guardrail that sits directly in front of package install commands and blocks risky packages before they execute.
attach-guard is a Claude Code plugin that intercepts package installation commands and evaluates them against policy before execution. It is not an advisory scanner. It is a hard enforcement boundary.
- Installs as a Claude Code plugin — no manual hook configuration needed
- Intercepts
npm install,pnpm add, directpip/pip3 install,python -m pip install/python3 -m pip installwrappers,uv pip install,go get/go install, andcargo add/cargo installcommands via PreToolUse hooks - Evaluates packages with the configured provider before install commands run
- Supports explicit provider configuration, including local Attach Open Score, Attach Open Score-compatible verdict endpoints, and legacy local BYO-token providers
- Denies known malware and high-confidence dangerous packages automatically
- Asks for confirmation on gray-band packages and provider-unavailable cases in local mode
- Rewrites unpinned installs to safe pinned versions when possible
- Logs every decision to a local JSONL audit trail
Attach Open Score is the first-party scoring direction. By default, open-score shells out to a local attach-open-score package command and can also be pointed at an explicit Attach Open Score-compatible HTTP verdict endpoint. No hosted/default Attach scoring endpoint is baked into attach-guard. See Attach Open Score provider semantics and Local Attach Open Score dogfood guide for public-safe verification.
Most security tools just say "no." attach-guard says "no, but here's a safe alternative."
When a risky version is blocked, attach-guard finds the newest version that passes policy and offers it as a replacement. Claude sees the safe alternative and can proceed immediately — your flow doesn't stop, it gets redirected to a safe path.
npm — axios v1.14.1 and v0.30.4 were publicly reported compromised versions published via a hijacked maintainer account:
> npm install axios
attach-guard evaluates:
axios@1.14.1 --> DENY (known compromised version)
axios@1.14.0 --> ALLOW (passes configured policy checks)
Result: ASK + rewritten command
"npm install axios@1.14.0"
pip — litellm v1.82.7 and v1.82.8 were malicious versions published to PyPI:
> pip install litellm
attach-guard evaluates:
litellm==1.82.8 --> DENY (compromised version)
litellm==1.82.6 --> ALLOW (passes all policy checks)
Result: ASK + rewritten command
"pip install litellm==1.82.6"
These examples illustrate the current enforcement flow: attach-guard blocks known compromised or policy-failing versions and offers a safe pinned alternative when one is available.
| Scenario | Example | Decision | What happens |
|---|---|---|---|
| Package is safe | npm install axios@1.14.0 |
Allow | Install proceeds normally |
| Pinned to compromised version | pip install litellm==1.82.8 |
Deny | Blocked — compromised version |
| Unpinned, latest is risky | npm install axios |
Ask + rewrite | Safe alternative offered: axios@1.14.0 |
| All versions fail | malware-only package | Deny | Blocked with clear explanation |
This works across all supported ecosystems — the rewrite uses the native pinning syntax for each:
| Ecosystem | Unpinned command | Rewritten command |
|---|---|---|
| npm / pnpm | npm install axios |
npm install axios@1.14.0 |
| pip | pip install litellm |
pip install litellm==1.82.6 |
| Go | go get golang.org/x/net |
go get golang.org/x/net@v0.25.0 |
| Cargo | cargo add serde |
cargo add serde@=1.0.200 |
Supported pip wrappers such as python -m pip install, python -I -m pip install, and uv pip install are evaluated before execution. If a wrapped pip command needs a safe-version rewrite, attach-guard asks for manual review instead of rewriting the wrapper command.
Your flow only fully stops when there is genuinely no safe version to offer.
attach-guard uses Claude Code hooks — not skills or MCP servers. The distinction matters:
- Hooks run automatically on every matching tool call. They enforce rules deterministically — Claude cannot skip or override them.
- Skills are instructions Claude follows when invoked. They guide behavior but cannot block actions.
- MCP servers provide advisory context. They inform but do not enforce.
Security enforcement requires interception at the tool-call boundary, before execution. Hooks are the only Claude Code extension point that guarantees this.
The fastest way to try the current packaged plugin. It installs the hook and local config. Attach Open Score is the first-party scoring direction, and open-score uses a local attach-open-score binary by default or an explicitly configured compatible verdict endpoint. No hosted/default Attach scoring endpoint ships in this package.
# Add the marketplace and install (one-time)
claude plugin marketplace add attach-dev/attach-guard
claude plugin install attach-guard@attach-devOr from within a Claude Code session:
/plugin marketplace add attach-dev/attach-guard
/plugin install attach-guard@attach-dev
Legacy BYO-token provider paths may prompt for a provider API token when explicitly configured. That token is for explicit local provider use only and is not required for the default Open Score path.
If the install/enable prompt didn't appear, re-trigger it with:
claude plugin disable attach-guard@attach-dev && claude plugin enable attach-guard@attach-devOr set the provider-specific token environment variable in your shell profile before starting Claude Code.
The prebuilt binary is downloaded automatically for your platform. The hook, config, and skill are all registered — no further setup needed.
Once running, the plugin provides:
- Automatic enforcement —
npm install,pnpm add, direct andpython -m pip/uv pipforms ofpip install,go get/go install, andcargo add/cargo installcommands are intercepted and checked /explain <package>— look up any package's risk score, alerts, and version history
If you want to develop or modify attach-guard, clone the repo and load the plugin directly. Requires Go 1.21+.
git clone https://github.com/attach-dev/attach-guard.git
cd attach-guard
claude --plugin-dir ./pluginThe binary auto-builds from source on the first /explain invocation.
Local claude --plugin-dir ./plugin development may not run the marketplace install/enable config flow. If Claude does not inject the plugin config in this mode, configure any required provider token in your shell before starting Claude Code.
For use without the plugin system, or to install the binary globally.
- Go 1.21+ (to build from source; not needed for the plugin install above)
- A provider API token only when using an explicit BYO-token local provider path
make buildMove the binary somewhere on your PATH:
# Option A: Move to a standard location
sudo mv attach-guard /usr/local/bin/
# Option B: Move to a user-local bin directory
mkdir -p ~/.local/bin
mv attach-guard ~/.local/bin/
# Make sure ~/.local/bin is in your PATH (add to ~/.bashrc or ~/.zshrc):
# export PATH="$HOME/.local/bin:$PATH"Verify it works:
attach-guard version
# attach-guard v0.1.0The default provider is local Attach Open Score:
provider:
kind: open-score
command: attach-open-scoreMake sure attach-open-score is on PATH, or set ATTACH_OPEN_SCORE_BIN=/path/to/attach-open-score. To use a local HTTP service instead, set provider.kind: open-score plus an endpoint.
attach-guard config init
# Default config written to ~/.attach-guard/config.yamlThis creates ~/.attach-guard/config.yaml with local Open Score defaults. Review Configuration before relying on provider behavior; no hosted/default Attach scoring endpoint is shipped in this package.
Add the following to your project's .claude/settings.json (shared with team) or .claude/settings.local.json (personal, gitignored):
{
"hooks": {
"PreToolUse": [
{
"matcher": "Bash",
"hooks": [
{
"type": "command",
"command": "attach-guard hook"
}
]
}
]
}
}For global protection across all projects, add it to ~/.claude/settings.json instead.
Try installing a known-compromised version to verify attach-guard blocks it:
> Install axios@1.14.1
Claude: I'll install axios@1.14.1.
[attach-guard] deny: axios@1.14.1: known compromised version
Then try a safe version:
> Install axios
Claude: I'll install axios.
[attach-guard] allow: package passes all policy checks
When Claude Code or Codex calls the Bash tool with a package install command (e.g., npm install axios, pip install requests, python -I -m pip install requests, go get golang.org/x/net, cargo add serde):
- The runtime fires the PreToolUse hook before execution
- The hook pipes the tool input JSON to
attach-guard hookvia stdin - attach-guard parses the command, evaluates packages against policy
- Returns a
hookSpecificOutputJSON response:permissionDecision: "allow"— install proceedspermissionDecision: "ask"— Claude shows the reason and asks the userpermissionDecision: "deny"— install is blocked, reason shown to the runtime
- On internal errors, exits with code 2 (blocking) to fail closed
Codex PreToolUse does not support permissionDecision: "ask" today. In the Codex adapter, Attach Guard maps manual-review ask decisions to deny so the package install does not proceed silently.
attach-guard evaluate <command> Evaluate a package manager command against policy
attach-guard hook [run|claude|codex]
Read runtime hook JSON from stdin and respond
attach-guard run [--dry-run] <claude|codex> [args...]
Run an agent after Attach Platform setup preflight and runtime hardening
attach-guard config init Write default config to ~/.attach-guard/config.yaml
attach-guard version Print version
attach-guard help Show help
# npm
attach-guard evaluate npm install axios
attach-guard evaluate npm install axios@1.14.1
# pip
attach-guard evaluate pip install litellm
attach-guard evaluate pip install litellm==1.82.8
attach-guard evaluate python -I -m pip install litellm
# Go
attach-guard evaluate go get golang.org/x/net
attach-guard evaluate go get golang.org/x/net@v0.25.0
# Cargo
attach-guard evaluate cargo add serde
attach-guard evaluate cargo add serde@=1.0.200
# Use as a Claude Code hook (reads JSON from stdin)
attach-guard hook
# Use as a Codex hook (reads JSON from stdin)
attach-guard hook codex
# Preview supported agent commands without starting the agent or checking setup
attach-guard run --dry-run claude
attach-guard run --dry-run codex
attach-guard run --dry-run claude --model sonnet
attach-guard run --dry-run codex --sandbox read-only "Review this diff"
# Run supported agents after `attach setup` has created ~/.attach/config.json
attach-guard run claude
attach-guard run codexattach-guard run applies runtime-native hardening before exec:
- Codex runs with
--sandbox workspace-write,--ask-for-approval on-request, and command network access disabled unless the user supplies an explicit supported setting. - Codex
danger-full-access,--yolo, and explicit command network enablement are rejected. - Codex receives an inline PreToolUse hook configuration that calls the current
attach-guard hook codexbinary unless the user already supplied hook config. The hook command path is resolved when the wrapped session starts; rebuilding or moving the binary affects new sessions, not already-running sessions. - Claude Code runs with session settings that enable the native Bash sandbox, fail closed if that sandbox is unavailable, select default permission mode, disable bypass/auto permission modes, and deny common direct web/secrets access rules.
- Claude Code
bypassPermissions,acceptEdits,auto,dontAsk,--dangerously-skip-permissions, and user-supplied--settingsare rejected by the wrapper. This uses Claude Code's native sandbox/settings surface; it is still not an Attach Platform or Gateway sandbox claim. - Claude Code receives
--plugin-dirfor the local/source Attach Guard plugin when the wrapper can findplugin/.claude-plugin/plugin.json, so a wrapped local run uses the fresh plugin instead of a stale marketplace cache. SetATTACH_GUARD_CLAUDE_PLUGIN_DIR=/path/to/pluginto choose a plugin explicitly, orATTACH_GUARD_CLAUDE_PLUGIN_DIR=offto skip this injection.
Default config location: ~/.attach-guard/config.yaml
attach-guard config init writes a local Open Score config. Attach Open Score is the first-party scoring direction, but no hosted/default Attach scoring endpoint is shipped in this package.
provider:
kind: open-score
command: attach-open-score # or set ATTACH_OPEN_SCORE_BIN
policy:
deny_known_malware: true
min_supply_chain_score: 70 # hard allow threshold
min_overall_score: 70
gray_band_min_supply_chain_score: 50 # hard deny below this
minimum_package_age_hours: 48 # deny versions newer than this
provider_unavailable_behavior:
local: ask # ask | deny | allow
ci: deny
unknown_behavior:
local: ask # Open Score UNKNOWN: ask | deny | allow
ci: deny
auto_rewrite_unpinned:
local: false # auto-pin to safe version?
ci: false
allowlist: [] # always allow these packages
denylist: [] # always deny these packages
package_managers:
npm: true
pnpm: true
pip: true
go: true
cargo: true
logging:
path: "~/.attach-guard/audit.jsonl"Use an Attach Open Score-compatible HTTP verdict endpoint explicitly:
provider:
kind: open-score
endpoint: http://127.0.0.1:8757/v0/verdict
timeout_seconds: 5The open-score provider posts only public package coordinates (ecosystem, name, version) and consumes verdict fields (decision, optional score, optional confidence, reasons, summarized source_refs). Provider failures and malformed responses map to UNKNOWN/provider-unavailable behavior, never ALLOW.
When endpoint is omitted, attach-guard shells out to the configured command and runs:
attach-open-score package --ecosystem <ecosystem> --name <name> --version <version>attach-open-score persists package-version verdicts in ~/.attach-open-score/scores.json by default. Set ATTACH_OPEN_SCORE_DB_PATH to choose a different cache file.
For BYO-token local provider configuration, see Advanced: BYO-token local providers.
ATTACH_GUARD_LOG_PATH— override log pathATTACH_GUARD_PROVIDER— override provider kindATTACH_OPEN_SCORE_BIN— use a specific localattach-open-scorebinary and forceprovider.kind: open-scoreATTACH_OPEN_SCORE_ENDPOINT— use a specific Open Score HTTP verdict endpoint and forceprovider.kind: open-score
Highest priority wins (later sources override earlier):
- Built-in defaults
- Plugin-bundled config (when installed as a plugin)
- User-global config (
~/.attach-guard/config.yaml) - Project-local config (
.attach-guard/config.yaml) - Environment variables
- Check allowlist/denylist
- Check provider availability
- Deny known malware
- Deny versions under minimum age (48 hours default)
- Use Attach Open Score-compatible provider verdicts when present (
ALLOW,ASK,DENY,UNKNOWN) - Deny scores below hard threshold (supply chain < 50) for legacy score providers
- Ask on gray-band scores (50-70)
- Ask on critical/high alerts
- Allow everything else
When you run an unpinned supported command such as npm install axios, pip install requests, python -m pip install requests, go get golang.org/x/net, or cargo add serde:
- attach-guard fetches candidate versions from the configured provider and evaluates them against policy
- If the latest passes policy, the command runs as-is
- If the latest fails but an older version passes, direct package-manager command shapes get a rewrite suggestion using ecosystem-native syntax
- In Claude Code mode, direct rewrites return
askwith the rewritten command viaupdatedInput - In Codex mode, manual-review
askdecisions are mapped todenybecause Codex PreToolUse does not supportasktoday - Wrapped pip commands are evaluated, but wrapper-preserving rewrites are not emitted yet; commands such as
python -m pip install requestsoruv pip install requestsask for manual review when a rewrite would be required - If no version passes, denies
Attach Open Score integration should be verdict-first: ALLOW → allow, ASK → ask, DENY → deny, and UNKNOWN → ask/warn locally by default. CI/team policy may map unknowns to deny by explicit configuration. See Attach Open Score provider semantics.
- Local/interactive mode: asks on provider failure
- CI mode: denies on provider failure (fail closed)
- Internal errors in hook mode: exit code 2 (blocks the install)
Every decision is logged to ~/.attach-guard/audit.jsonl:
{
"timestamp": "2026-01-15T10:30:00Z",
"user": "dev",
"cwd": "/home/dev/project",
"package_manager": "npm",
"original_command": "npm install example-malware@1.0.0",
"decision": "deny",
"reason": "example-malware@1.0.0: known malware alert",
"packages": [{"ecosystem":"npm","name":"example-malware","selected_version":"1.0.0","score":{"supply_chain":0,"overall":0},"provider_verdict":{"decision":"DENY","risk_score":95,"reasons":["known-malware-synthetic"],"source_refs":["osv:GHSA-0000-0000-0000"]},"alerts":[{"severity":"critical","title":"known malware","category":"malware"}]}],
"provider": "mock",
"mode": "claude"
}New setup should use Attach Open Score, either through the hosted endpoint with
ATTACH_OPEN_SCORE_API_TOKEN or through the local attach-open-score command.
Legacy provider adapters may remain in code for existing private workflows, but
Socket data must not be used as default Attach scoring or as an Attach Open
Score input/source unless a future written partnership and policy explicitly
permit it.
Provider calls can fail when the hosted service or local scorer is unavailable,
rate-limited, or out of quota. Provider startup/unavailability follows
provider_unavailable_behavior; local defaults to ask/warn, and CI/team
fail-closed behavior requires explicit configuration such as
provider_unavailable_behavior.ci: deny. Individual scoring errors can still
block a pinned install, and unpinned rewrites need real provider results to
identify a version that passes policy.
- Direct
pip/pip3,python -m pip/python3 -m pipwrappers with safe interpreter flags before-m,uv pip,go get/go install, andcargo add/cargo installare supported - Rewrites are emitted only for direct package-manager command shapes; wrapped pip commands are evaluated but ask for manual review when a safe-version rewrite is needed
uv add/uv sync, pip extras/range/constraint-aware handling, Cargo requirement syntax, and non-semver Go queries are intentionally deferred to manual review rather than being auto-evaluated- Default hosted scoring is not baked into this package; local Attach Open Score runs through
attach-open-score, and HTTP scoring requires explicitopen-scoreendpoint configuration or platform-issued credentials fromattach setup - Local BYO-token providers can be affected by upstream availability, rate limits, or quota during evaluation and version rewrite
- No transitive dependency analysis
- No lockfile graph support
- Single provider at a time
- No org-level policy distribution
- No remote audit export
- Vendor API response formats may vary for local BYO-token providers; adapters are based on documented endpoints
# Run all tests
make test
# Build
make build
# Evaluate a command
./attach-guard evaluate npm install lodashCross-compile plugin binaries for all platforms:
make plugin-buildTest the plugin locally:
claude --plugin-dir ./pluginMIT