Tens of thousands of instances of a popular agentic AI platform sit exposed on the public internet. Hundreds of advisories have been written against it in the months since launch, and its skill marketplace carries roughly one in five entries that are malware. The architectural fault is ambient credential authority in a single process with no audit log.
Ambient authority
The pattern is familiar. One process holds every credential you handed it. The LLM driving that process can call any tool it has been wired to. OAuth scopes were designed for a human at a keyboard, so they are too coarse to constrain an agent operating at machine speed across the full set of scopes you granted. When that process is compromised, or when the model is talked into something you did not intend, the blast radius is everything you can do on the systems those tokens reach.
Per-process credential sprawl is the symptom. A single node process holding the keys to your mail, your chat, your files, and your source control is the failure mode.
Treating it as a configuration problem inside that same process produces a more configurable single point of failure, not a smaller one.
What we built instead
We built wirken so you can break your agent's authority into pieces and audit what it does with each piece. Four mechanisms carry the design.
The tier of every tool call is computed in Rust from the tool name and its arguments. A model that emits {"tool": "exec", "command": "rm -rf /"} resolves to Action::ShellExec for a command that is not on the inspection allowlist, which is Tier 3, which asks you before it runs. The model does not choose the tier.
Each channel adapter runs as its own OS process with its own Ed25519 identity and is handed only its own platform credentials. The gateway checks the channel on every inbound frame against the one the connection authenticated as. Adapters run as the same operating-system user as the gateway.
Credentials sit in an XChaCha20-Poly1305 vault whose key is a passphrase-wrapped file in the data directory. Decrypted secrets are wrapped in VaultSecret, which implements none of Display, Debug, Serialize or Clone and whose memory is zeroed on drop, so a println!("token: {token}") does not compile.
Every session writes a SHA-256 hash-chained event log with chain heads signed by the installation's Ed25519 audit key. wirken audit verify checks the chain and the head signatures offline; wirken sessions verify re-checks message hashes and re-runs read_file and list_files.
What this project deliberately does not do
The LLM is not in the gating path. The tier is decided in Rust from the tool name and arguments before the call reaches the tool. The model proposes; the gateway decides; the audit log records.
We do not host workflows in the switchboard. Workflows belong in skills. A skill can implement multi-agent shapes internally, including planner-executor splits, voting, and self-critique. The switchboard routes; the routed surface is the skill. Hosting an FSM in the switchboard would widen the trusted base for every workflow that anyone anywhere wants to run, and the trusted base is exactly what wirken keeps small.
The name
The original Gebrüder Ottenheimer was a cotton and linen weaving mill in Württemberg, founded in 1862 by Salomon and Joseph Ottenheimer. The mill wove isolated threads into a usable fabric. This project does the same with isolated communication channels. The rest of that history, including the forced sale in 1937 and what the name commits the project to remember, sits at the historical brief.
Your agents or operators will hold the credentials this decade. We built wirken so that stays your choice, with enough evidence in the audit log to answer for it after the fact.