ClawShell π‘οΈ

Powered by Runta. The essential safety harness for OpenClaw/Hermes Agent's PII & Sensitive Credentials.
π Introduction
ClawShell is a security-privileged process for the OpenClaw/Hermes Agent ecosystem. It sits between OpenClaw/Hermes Agent and upstream LLM API providers (OpenAI, Anthropic, OpenRouter), performing virtual-to-real API key mapping and DLP (Data Loss Prevention) scanning on request and response bodies. It can also expose an Email read endpoint with sender allowlist/denylist filtering.
OpenClaw/Hermes Agent never holds real API keys, only virtual keys that ClawShell swaps for real ones before forwarding requests upstream. Real keys are stored in a privileged config directory (/etc/clawshell) protected by Unix file system permissions.
Key Features
1. API Token Secure Binding
ClawShell maps virtual API keys to real provider keys so that OpenClaw/Hermes Agent never has direct access to real credentials.
- Key Isolation: Real API keys are stored in
/etc/clawshell/clawshell.toml, readable only by theclawshellsystem user. OpenClaw/Hermes Agent holds only virtual keys. - Multi-Provider Support: Maps keys to OpenAI or Anthropic, injecting the correct authentication header format (
Authorization: Bearerfor OpenAI,x-api-keyfor Anthropic).
2. PII Safety Net (DLP)
ClawShell scans HTTP request and response bodies for sensitive data using configurable regex patterns.
- Request Scanning: Detects PII (SSNs, credit card numbers, emails, etc.) in outbound requests. Patterns can be configured to either block the request or redact the matched text before forwarding.
- Response Scanning: Optionally scans upstream responses and redacts detected PII before returning to OpenClaw/Hermes Agent. Streaming (SSE) responses are passed through without scanning.
- Custom Patterns: Define sensitive data patterns using regex in the TOML config, each with a
blockorredactaction.
3. Sensitive Email Isolation
ClawShell supports sender-based email filtering so each virtual key only sees mailbox content based on sender rules.
- Sender Filtering: Filter emails by sender.
- Key Isolation: IMAP credentials are stored in
/etc/clawshell/clawshell.toml, readable only by theclawshellsystem user. OpenClaw/Hermes Agent holds only virtual keys. - Provider Support: Built-in Gmail and Outlook presets, with manual IMAP setup for other providers.
4. OAuth Authentication (Codex / ChatGPT)
ClawShell supports OAuth-based authentication as an alternative to static API keys.
- Device Code Flow: Log in via a one-time code β no browser required on the server. The onboard wizard prints a URL and code; authorize from any device.
- Automatic Token Refresh: Access tokens are refreshed transparently before they expire.
- Request Translation: Automatically translates OpenAI Chat Completions API requests to the ChatGPT Responses API format when using Codex OAuth.
5. Runtime Statistics
ClawShell exposes running counters at GET /admin/stats so operators can audit proxy activity since startup and across restarts.
- What's Counted: Total requests served, total upstream
prompt_tokens/completion_tokens/total_tokens(parsed from non-streaming JSON responses β SSE streams are not counted), number of emails hidden by the sender policy, and a per-address breakdown of filtered senders. - Loopback-Only: The endpoint is reachable without a virtual key but only from
127.0.0.1/::1peers; non-loopback clients receive403. - Persistent: Counters are flushed to disk every 30 seconds and on graceful shutdown. The location is a required config field β set
[stats] persist_path = "..."inclawshell.toml(typically/var/lib/clawshell/stats.jsonunder the hardened systemd unit, since/etc/clawshellis read-only there). - Bounded: The filtered-address map is capped at 10,000 unique entries; further unique addresses are aggregated under an
<overflow>bucket so memory stays bounded.
6. Seamless Integration
- Drop-in Sidecar: The
clawshell onboardwizard configures exactly one downstream LLM client per run β either OpenClaw or Hermes Agent β to route all requests through ClawShell's proxy. See Agent Target (pick one). - No External Dependencies: Uses Unix file system permissions to protect secrets. No IdP, Vault, or external key management service required.
7. Ultra Lightweight and Scalable
- Runs in under 10MB of memory.
- Written in Rust with Tokio.
Architecture
β security boundary (Unix File System Permissions)
β
β βββββββββββββββββββββββββββββ
β β /etc/clawshell β
β β β real API keys β
β β β DLP patterns β
β β β email sender rules β
β β β IMAP account creds β
β ββββββββββββ¬βββββββββββββββββ
β reads β
β ββββββββββββ΄βββββββββββββββββ
ββββββββββββββββ REQUEST β β β REQUEST ββββββββββββββ
β βββ(virtualββββ«ββΊβ ClawShell βββ-(real key,ββββΊβ β
β OpenClaw or β key) β β β PII redacted) β OpenAI / β
β Hermes Agent β β β DLP scan β β Anthropic/ β
β holds only β RESPONSE β β real-key mapping β RESPONSE β OpenRouter β
β virtual keys ββββββββββββββββββ€ email sender filtering βββββββββββββββββββ€ β
β β β β β ββββββββββββββ
β β EMAIL GET β β β IMAP fetch ββββββββββββββ
β ββββ(virtualβββ | |βββ(real key)ββββΊ| |
β β key) β β β β IMAP β
β β β β β | Provider β
β β β β β RESPONSE β Gmail / β
β β filtered β β βββββββββββββββββββ Outlook / β
β βββββββββββββββββ| | β custom β
ββββββββββββββββ β | | ββββββββββββββ
β βββββββββββββββββββββββββββββ
OpenClaw/Hermes Agent only holds virtual keys and cannot access the real API keys stored in the privileged config.
ClawShell swaps virtual keys for real ones and scans for PII before forwarding requests upstream.
ClawShell also enforces sender-based filtering before returning email data.
Installation
Cargo
# Requires privilege to set up the security boundary
NPM
# Requires privilege to set up the security boundary
Build from Source
RUSTFLAGS="--remap-path-prefix==. --remap-path-prefix==/"
Cross-compile on Linux/arm64
CARGO_TARGET_X86_64_UNKNOWN_LINUX_MUSL_LINKER="/tmp/x86_64-linux-musl-cross/bin/x86_64-linux-musl-gcc" \
RUSTFLAGS="--remap-path-prefix==. --remap-path-prefix==/" \
Advanced Usage
Onboarding
The onboard command is an interactive setup wizard that must be run with sudo. It:
- Asks which downstream agent to wire through ClawShell β OpenClaw or Hermes Agent (exactly one per run).
- Creates the
clawshellsystem user. - Creates and secures
/etc/clawshell(mode 700) and/var/log/clawshell. - Walks you through provider selection, API key entry, and virtual key generation.
- Writes the ClawShell config to
/etc/clawshell/clawshell.toml. - Wires the chosen agent through ClawShell (patches
~/.openclaw/openclaw.jsonfor OpenClaw, or runshermes config setfor Hermes). - Starts the ClawShell daemon.
More Commands
# Start (daemonizes by default)
# Start in the foreground
# Start with a custom config file
# Check status
# View logs
# Restart / Stop
# Migrate config schema to current version
By default ClawShell listens on 127.0.0.1:18790.
You can override the bind address at runtime with environment variables:
CLAWSHELL_SERVER_HOST=0.0.0.0 CLAWSHELL_SERVER_PORT=17890
Customized Configuration
ClawShell reads its config from /etc/clawshell/clawshell.toml. You can view or edit it with:
A minimal config looks like this:
= "0.2.0"
= "info"
[]
= "127.0.0.1"
= 18790
[]
= "https://api.openai.com"
= "https://api.anthropic.com"
# Virtual-to-real API key mappings
[[]]
= "vk-alice-001"
= "sk-your-real-openai-key-here"
= "openai"
[[]]
= "vk-claude-001"
= "sk-ant-your-real-anthropic-key-here"
= "anthropic"
# Data Loss Prevention (DLP)
# action = "block" -> reject the request with 400
# action = "redact" -> replace matches with [REDACTED:<name>] and forward
[]
= false
= [
{ = "ssn", = '\b\d{3}-\d{2}-\d{4}\b', = "redact" },
{ = "visa_card", = '\b4[0-9]{12}(?:[0-9]{3})?\b', = "redact" },
{ = "amex_card", = '\b3[47][0-9]{13}\b', = "redact" },
]
# Email secure endpoint
[]
= true
= "allowlist"
= ["alice@example.com", "@trusted.org"]
= []
= 50
[[]]
= "vk-email-001"
= "bot@gmail.com"
= "abcd efgh ijkl mnop"
= "imap.gmail.com"
= 993
# Outlook preset example:
# imap_host = "imap-mail.outlook.com"
OAuth Authentication (Codex / ChatGPT)
Instead of a static API key, you can authenticate via OAuth using your ChatGPT / Codex account.
Setup via onboard
During sudo clawshell onboard, select "Codex / ChatGPT (OAuth)" as the provider. The wizard will start a device code flow:
- A URL and one-time code are printed to the terminal.
- Open the URL on any device and enter the code.
- Once authorized, tokens are saved automatically to
/etc/clawshell/oauth/.
No browser is required on the server.
Manual configuration
To configure OAuth manually in clawshell.toml:
# OAuth-backed key β no real_key needed
[[]]
= "vk-codex-001"
= "oauth"
= "codex"
= "openai"
# OAuth provider definition
[[]]
= "codex"
# Optional overrides (defaults work for ChatGPT):
# client_id = "app_EMoamEEZ73f0CkXaXp7hrann"
# auth_url = "https://auth.openai.com/authorize"
# token_url = "https://auth.openai.com/oauth/token"
ClawShell handles token refresh automatically. When using Codex OAuth, requests to /v1/chat/completions are translated to the ChatGPT Responses API format and routed to chatgpt.com/backend-api/codex.
If start, restart, stop, config --edit, onboard, or uninstall reports that migration is required, run:
See clawshell.example.toml for a full example.
Agent Target (pick one)
sudo clawshell onboard begins with a single, mandatory choice:
=== Agent Target ===
? Which downstream agent should ClawShell wire through?
> OpenClaw
Hermes Agent
Each onboard run configures exactly one downstream client. There's no "also configure the other one" path β switching later means re-running sudo clawshell onboard and picking the other target. The prompt has no default preselection, so you pick explicitly every time.
OpenClaw target
When you pick OpenClaw, the wizard:
- Backs up
~/.openclaw/openclaw.json(numbered.bakfiles, mode 000). - Shells out to
openclaw config setto patch three paths:env.CLAWSHELL_API_KEY,agents.defaults.models.clawshell/<model>, andmodels.providers.clawshell. - Writes a
get-email-messagesskill bundle to<openclaw_root>/skills/when email integration is enabled. - Offers to run
openclaw models set clawshellandopenclaw gateway restartat the end.
This is the historical onboarding flow and is unchanged by the target-selection rework.
Hermes Agent target
When you pick Hermes Agent, the wizard:
-
Skips every OpenClaw step β
~/.openclaw/is not touched. -
Writes a
get-email-messagesskill bundle to~/.hermes/skills/(owned by your invoking user, not root) when email integration is enabled. Hermes auto-discovers skills from that directory. -
Shells out to
hermes config setto write:Key Value model.providercustommodel.base_urlhttp://<server_host>:<server_port>/v1model.defaultthe model ID you chose during onboard model.api_keyyour ClawShell virtual key (never the real upstream key)
The hermes binary must be on your PATH. ClawShell drops root privileges before invoking it so writes land under your normal user account, not root's.
Manual Hermes configuration
If you'd rather skip the wizard's Hermes integration, run the equivalent commands from your user account (not root):
Then verify with hermes config show.
Reverting Hermes
Hermes has no config unset subcommand. To detach Hermes from ClawShell, set the provider back to auto-detect and Hermes will pick another upstream based on the credentials it still has:
Uninstall
License
This project is licensed under the Apache License 2.0.