The threat
An agent that acts for you holds your credentials, reads input from sources you do not control, and runs commands with whatever access you gave it. What makes it useful is what makes it dangerous. It acts without asking. A hidden instruction in a chat message, an old comment in a shared document, or an email you never opened is enough to turn a helpful assistant into one that leaks your password or changes a file you forgot it could touch.
Wirken puts a gate between the agent and what it can do. Reading and writing files in the agent's workspace runs free. Twenty-three inspection commands ask unless you have recorded a grant. Every other command and every MCP tool call asks each time. Every decision is recorded so you can go back later and see what happened. What sits outside these controls is listed in the threat model.
Install
Linux and macOS:
curl -fsSL https://wirken.ai/install.sh | sh
Windows: docs/windows.md
Verify the installer first? See /wirken/install.html.
What's in the box
Ten channels: Signal, WhatsApp, Telegram, Matrix, Discord, Slack, Teams, Google Chat, iMessage, and a web chat. Sixteen skills your agent can use, from coding tools to notes and weather. One agent, one sealed log, one vault. Each channel runs separately and holds only its own credentials, so if Slack is compromised, your Signal credentials stay safe.
The switchboard
Every step verifies who is talking. Channels reach the agent only after they identify themselves. You can run one agent for everything, or a different agent for each channel, each with its own model and credentials. The agent only touches the outside world through tools you approved. The log of everything that happened is sealed with a signature only your install can produce.
Policy is text
Each skill is a plain text file that starts with a few lines of rules. The git skill, in full:
---
name: git
description: Git version control operations
permissions:
tools:
allow: [exec]
egress:
mode: deny
inference:
allow: ["*"]
---
Eight lines you can read at a glance. This skill can run commands on your computer but cannot reach the internet, and can use any AI model. No special language to learn, no separate config file. The file itself is the rule.
The audit log
Every tool call is recorded in a hash-chained log before it runs, and its result after. You can check the whole log without an internet connection:
$ wirken audit verify
WARNING (audit anchor): report-only: the hash chain and self-attested chain-head signatures were checked, but no operator trust anchor was consulted, so a same-UID rewrite is not detected. Pass --require-signed with an out-of-band --anchor for tamper-evident verification.
Audit log integrity: OK
16 rows verified across 2 sessions, hash chain intact.
Chain-head signatures: 4 verified.
Signing key ids seen: f7aa9d699e173fa6857e355a9abef44672aff79deded7260362ceaab66a2388b
If anyone changes an entry, the hashes after it no longer match and the check fails. The chain heads are signed with a key your installation holds. Checking against a copy of that key kept off the machine is what detects a rewrite by someone with access to the installation, which is what the warning in the output says.