An assistant with your key, and your finger on the button.
Themis explains an answer, writes a request's script or its checks, imports a spec from a URL, builds a mock, builds a whole app — through Sonda's own validated code, never by running text the model produced. Sending a request, deleting anything or running a folder becomes a button in the chat that you press, or do not.
At a glance
The model proposes. Sonda's own code does it. You press the button.
Themis talks to the provider you pick — Anthropic, OpenAI, xAI, or any OpenAI-compatible local server — with your key, sealed under the account password and never handed to the window. The provider addresses are fixed in the code; only the local one can be changed. It sees the question, the last twelve turns and, if you leave the box ticked, the active tab: method, URL, headers with secrets hidden, the body cut short, the last answer, the test results.
What comes back is a tool call, and a tool call runs the same validated function a click would, under the workspace lock, with the member's permissions checked. The model's words never become code. Sending, deleting and running a folder wait in the chat for your click — and the conversation lives in memory while Sonda runs: lock or quit and it is gone.
What leaves this computerForty-two tools, all of them Sonda's own code.
A tool call runs the same validated function a click would, under the workspace lock, with the member's permissions checked. The model's words never become code.
Reads and explains.
Explain this answer. Why is this broker pane amber. What does this flow do. Read a device's last values. List the projects, the flows, the Echoes, the apps.
Builds from the document, not from a guess.
A project from a documentation URL — it fetches the page, finds the OpenAPI, WSDL, OData or FHIR document behind it, and imports it rather than guessing. Folders, requests with any body type, variables, auth told to use placeholders and never real values, checks from the last answer, a script into the editor.
Mocks and apps, tried before they are handed over.
An Echo with a realistic data set and endpoints answering by data, body or script. An app with containers, screens, a dashboard, watchers and functions — tried before they are handed over. An On watch from a device to a request.
Proposes, and waits.
Send this request. Delete this. Run this folder. Each one is a card with No and the verb, and Show me to open it on screen first — on all but a delete. The outcome comes back as a turn. A delete re-checks the permission at the moment you press.
Refuses what is not its to do.
Frozen, locked and paused items. Anything the member's flags do not allow. Submitting, approving, locking, pausing and merging — never Themis's. Off for app runners and while Sonda is locked.
Guards against what it reads.
Every tool result and every screen block is sealed as untrusted data with a random nonce, and the system prompt says not to follow instructions inside it. About thirty injection phrases are removed visibly, with a line that says so. Web reads only from hosts named in the conversation, GET only, 16 MB.
What Themis does without a click, and what she never does.
Said exactly, because the difference matters in a regulated environment. Building is Themis's: a project, a mock, an app, a check, each step shown as it goes. Sending, deleting and running a folder are yours: a card in the chat, with the verb on the button.
Beside it, one turn: a question with the request on screen, two changes made without asking, and a send that waits for the person.