Security
The boundary is your OS user
Wisp trusts every process running as the same operating-system user as the Runtime. That is the same line Chrome, Docker Desktop and password managers draw: if something hostile already runs as you, the Runtime's token is not what stands between it and your data.
The token, the loopback listener, and the Host and CORS checks protect against
everything that is not your user:
- a web page in your browser calling
http://127.0.0.1:3008/v1/…; - other accounts on the same machine;
- anything on the network. By default the Runtime listens on
127.0.0.1only. Ahostother than loopback works only together withallowedHosts, and the token must then reach it over TLS: see Reaching it from another machine.
Protecting against your own user (checking which program is calling, guarding the Project folder or the binary) is outside what Wisp does.
The token and the lock file
The token is in runtime.json, which the installer creates readable only by you. The
running Runtime writes its address and token to runtime.lock, also readable only by you,
so that the CLI and adapters can find it. That gives nothing new to your own processes,
which could already read the Runtime's environment, and nothing to other users.
On Windows there are no Unix file modes; both files sit under your user profile, which only you can read.
There are four kinds of token:
| Token | What it can do |
|---|---|
the one token (agentToken) | everything Wisp Desktop's chat can do — run prompts, change Rules, manage connectors, read the audit — and create, list and revoke the named tokens |
a named full token | the same, except the named tokens themselves |
a named run token | only what a Channel adapter needs: open Sessions on the Desk, start Runs in them, read every Session and Run, cancel any Session's Run, answer the Prompts of Runs it started (approve or deny only), deliver to its own Channel, and install the speech model |
| a Host Agent Credential | only the Runtime's /mcp endpoint, which Hermes hands its tasks to; every other route is 403 |
No token can stop the Runtime: there is no stop endpoint.
Treat a leaked token like a leaked password. A named one: wisp tokens revoke <name>, and
the others keep working. The one token: change agentToken in runtime.json and restart.
See A token per device.
Nobody approves anything mid-Run, unless you allow it
A Run started from the CLI or /v1 runs under the policy confirmation mode unless it asks
for another: any tool whose decision is ask is refused, and the agent continues without
it. Only tools a Rule explicitly allows can act. A caller may ask for the interactive mode
instead, which confirmationModes allows by default: the Run then parks on an ask, and
whoever holds the token and the Run's id can approve it. Set confirmationModes to
["policy"] to refuse that, which also stops the chat bots.
The consequence: if a web page, a connector result or a document tries to get the agent to
"send this", the default result is a refusal, not an approval card that nobody sees.
That puts all the weight on your Rules. See Permissions, including the
warnings about bash_run.
Turn the bash_run sandbox on wherever commands run without you reading
each one first; Sandbox says when. Without it (wisp sandbox off, the
default) a bash_run command runs as the Runtime's own user and can read the key source,
and the file tools may reach outside the Project, each such call on a card. Sandbox or not,
the file tools refuse Wisp's data directory, .env files, /etc and the usual secrets in
your home folder; Tools lists them.
What reaches the model
Every connector result goes through one filter that removes control characters, invisible and bidirectional characters, and Wisp's own framing tags, so a tool's output cannot pretend to be part of the conversation's structure. What the filter cannot tell is who wrote a line; see Who wrote what.
An approval request names the tool and its arguments. The arguments are what the model wrote, a drafted message among them, so read them before you approve.
The model is reached through the Wisp proxy, which runs inside a trusted execution environment, and the Runtime sends the proxy nothing unless that environment is verified: no prompt, no web search, no Gmail, Calendar or Drive call. Traffic that does not go through the proxy is not gated: your own connectors, the Telegram connector, sign-in and the update check. See Attestation.
Secrets
- On disk: the stores are encrypted with ChaCha20-Poly1305. The key is kept in the OS keychain, or, on a server, sealed in the data directory under a key file kept outside it. Connector secrets live in that encrypted store, never in a Project folder.
- The audit keeps each tool call's arguments and result, cut short and with values
under secret-looking names such as
tokenorpasswordreplaced by[REDACTED]. A text a tool sent is in it. - On the way in:
--header-stdinand--api-hash-stdinread a secret from the terminal, because a command line can be seen by other users on the machine. runtime.json: a file owned by another user is refused (exit 78), and so is a file holding a secret that others can read:chmod 600it.telegram.jsonholds the bot's token and itsruntoken oncewisp telegram setuphas run, and is written readable only by you.slack.jsonholds the app's bot and app-level tokens and itsruntoken oncewisp slack setuphas run, and is written readable only by you.
Not available yet
The Runtime does not terminate TLS itself. To reach it from another machine, use
Wisp Desktop, which opens an SSH tunnel itself, put a
reverse proxy with TLS in front of it, or tunnel to its loopback port (for example with
ssh -L).