Private AI agent that
lives anywhere
Run Wisp from the terminal, put it on a schedule, or message it on Telegram and Slack. Nobody, including us, can read, store or train on your data.
Channels
One agent. 3 ways to reach it.
Telegram and Slack
Message your agent like a teammate. It answers from your files and connectors, wherever you are.
Telegram and Slack carry the messages you trade with the bot, so they can see them. The model still only runs inside the verified enclave.
How it works
Private by design
Built with cryptographic proof that nobody can read, store or train on your data.
Wisp Kernel
Hard limits on what any tool call can do. Stops the agent from misbehaving.
Sealed Hardware (TEE)
Wisp Anonymizer
Runs web searches for you, so Google and other sites can’t profile you.
LLM in sealed hardware
AI runs in a Trusted Execution Environment (TEE). Nobody, not even us, can see inside.
Proof
Lives on your machine.
Thinks in a sealed box.
The agent, its sessions and your credentials stay on the machine you install it on. The LLM runs in sealed hardware: before anything is sent to AI, Wisp checks a hardware-signed receipt proving what code is inside.
No receipt = nothing leaves your machine.
Live cryptographic attestation
Checked by the Runtime before a prompt, a web search or a Gmail, Calendar or Drive call goes out. You can verify the same values yourself.
- App IDmatches expected
- 366d10efa0f7d2af57676742b79af3c069d730c9
- Compose hashmatches expected
- 5b0859c6af75672831eccd6ac6f98082a0b874c82fa273a01b6816c595e43dd2
- Image digestmatches expected
- e96833f976d33103fd1cfc0224544f5e1370c217d9a2f5c57c31e27da786db38
- MRTDmatches expected
- f06dfda6dce1cf904d4e2bab1dc370634cf95cefa2ceb2de2eee127c9382698090d7a4a13e14c536ec6c9c3c8fa87077
Permissions
A rule for every tool
Allow, ask or deny per tool. Unattended runs refuse anything you haven’t allowed.
Storage
Encrypted on your disk
Sessions and connector secrets are encrypted on the disk. Wisp keeps no copy.
Network
Loopback only
The API listens on 127.0.0.1 behind a token. Nothing is exposed unless you tunnel in.
Audit
Every tool call on record
wisp auditlists every tool call, secrets redacted. The log stays with you.
Use cases
How teams use Wisp Runtime
Always-on watchdog
Runs 24/7 on your own server
Scans logs for updates or breaches, and alerts on a schedule
Alerts the right people, within the rules you set
Single source of truth
Connect Wisp to your knowledge base, docs and team chat
Anyone can @Wisp and get the answer from the relevant source
Prompts are processed in a sealed enclave and never used for training
Install
From zero to first answer in about five minutes
One script installs the wisp command and a background service that survives logouts and restarts.
- macOS on Apple silicon or Intel, Linux on x86-64 or arm64
- Signed releases, verified by the installer before anything changes
- No GPU needed: the model runs in the enclave, not on your box
Install macOS or Linux
curl -fsSL https://usewisp.io/runtime/install.sh | shSign in on a server: wisp login --manual
wisp loginAdd balance terminal runs use a prepaid balance
wisp billing topup 10Ask something the current folder becomes the project
wisp run "what does this project do?"
Compare
Wisp Runtime vs Hermes and OpenClaw
The same always-on agent idea. The difference is who can read what it works on.
| HermesNous Research | OpenClawopen-source agent | ||
|---|---|---|---|
| Agent that lives anywhere | |||
| AI provider can’t read your prompts | local LLMs only | local LLMs only | |
| Never used to train models | provider policy | provider policy | |
| Hardware proof before anything is sent | |||
| Connector keys and chats encrypted on disk |
Testimonials
What our users run on it
“a7f3 9d2e 4c81 b6f0 e2c4 8d1a 5b9c”
“xK9m R2pL qT7v N4Zb H8wF j5Yc”
We literally don’t know.
FAQ
Questions, answered
It is the same Wisp agent without the app window: one program,
wisp, that runs as a background service and takes prompts from your terminal, a schedule, a chat bot or a local HTTP API.If Wisp Desktop is open on the same machine, the CLI uses the Runtime inside the app, and the two share sessions and connectors.
On macOS (Apple silicon or Intel) and Linux (x86-64 or arm64), on your laptop or on a server, as a launchd agent or a systemd user service. The installer does not support Windows yet.
No. The agent, its sessions and your credentials stay on your machine, but the model runs in a trusted execution environment, so you need no GPU. Before anything is sent to it, the Runtime checks that the environment is genuine.
No. Prompts only leave your machine for a verified enclave that nobody, including us, can look into. Sessions and connector secrets are encrypted on your disk, with the key kept in your OS keychain.
Attestation gates everything that goes through the Wisp proxy: prompts, web search and the Gmail, Calendar and Drive connectors. Your own connectors, the Telegram connector, sign-in and the update check go out directly, and the Runtime trusts every process running as your OS user.
Only what a rule explicitly allows. A tool set to ask is refused in an unattended run and the agent carries on without it, so a web page that tells it to “send this” gets a refusal, not an approval nobody sees. Every refused call is in
wisp audit.The Runtime listens on 127.0.0.1 only. Wisp Desktop reaches a server over SSH without opening a port; anything else goes through a reverse proxy that terminates TLS, with a named token for each device.
Terminal runs are paid from a prepaid balance on your Wisp account, which you top up with
wisp billing topup. It does not expire, and it pays for a Runtime on any machine signed in to the same account.