> ## Documentation Index
> Fetch the complete documentation index at: https://docs.zhentan.me/llms.txt
> Use this file to discover all available pages before exploring further.

# Self-Hosting the Agent

> Where the self-hostable agent stands today, and what's coming.

Zhentan's long-term goal is that **you run your own agent** — the service that screens your transactions and holds the co-signing key answers to you, on your machine, while the Zhentan backend keeps doing what a backend should: canonical state, the Safe Transaction Service mirror, and gas relay.

The architecture for this already exists. This page is honest about which parts are real today and which are on the roadmap.

## What Exists Today

The agent runtime is a **fully separable process** with a hard boundary:

* It talks to the backend exclusively over an authenticated HTTPS API — configure `RUNTIME_API_URL` and it can run anywhere
* It has **no database credential** (enforced by lint, dependency guards, and a boot test) and accepts no inbound connections
* It holds the signing key (`AGENT_PRIVATE_KEY`) behind a verifying authority: it recomputes every transaction hash, reads the Safe's owners and nonce from chain (`RUNTIME_RPC_URL`), and signs only what **its own decision log** proves it screened
* It ships with deployment tooling: `./runtime/run.sh build && ./runtime/run.sh start` (PM2), a localhost health endpoint, and an env template at `runtime/.env.example`

```
runtime/.env
├── RUNTIME_API_URL      # the backend — anywhere reachable
├── RUNTIME_API_TOKEN    # authenticates the runtime to the backend
├── AGENT_PRIVATE_KEY    # the ONLY signing key in the system
└── RUNTIME_RPC_URL      # chain reads for pre-sign verification
```

If the runtime is offline, screened transactions **wait** — nothing executes without a verdict, and your relay-only and backup-key flows continue untouched.

## What's Coming

Today there is one shared agent identity, so running the runtime yourself is an operational split, not yet *your* agent. The remaining milestones:

| Milestone                       | What it adds                                                                                                          |
| ------------------------------- | --------------------------------------------------------------------------------------------------------------------- |
| **Agent identities & pairing**  | Each Safe maps to a named agent identity; you pair your runtime to your wallet                                        |
| **Per-agent credentials**       | Your runtime's credential is scoped to your Safe alone — it cannot lease anyone else's work                           |
| **Policy snapshots & evidence** | Your behavioral profile flows down to your runtime; screening evidence flows back up                                  |
| **Your own signer key**         | A per-user agent key, swapped in as your Safe's owner via a reviewed owner-management transaction                     |
| **Your agent talks to you**     | Review prompts come from your own runtime through your own channel, and your approvals return as first-party evidence |

The signing rules, wire protocol, and verification model shipping today were built for that end state — self-hosting swaps credentials and configuration, not architecture.

<Note>
  Until per-agent credentials ship, the runtime deployment is operated by Zhentan. The security properties above — user-held key majority, screened-only signing, the Safe-app override — hold either way.
</Note>