Skip to main content

The Wisp agent

The Wisp agent is the local orchestrator that runs on your machine as a sidecar process. Everything the LLM needs or wants to do routes through the agent, and nothing touches your filesystem or the network without its say-so.

note

The Wisp agent is the Wisp Runtime running inside the desktop app. The same Runtime is what wisp serve starts when you use Wisp from a terminal.

The agent sends only the minimal context the LLM needs to reason, and the response comes back through the same encrypted path. Everything else (reading, writing, execution) stays local.

Confirmation before risky actions​

Tools whose Rule resolves to ask (by default writing files, shell commands, changing a scheduled job, every connector tool and a few more; see Tools) stop before they run: the Kernel holds the call and the desktop app shows a card with the exact action. You can approve it once, deny it, or allow the tool for this Project or everywhere.

Silence is never consent. A confirmation nobody answers is denied after confirmTimeoutMs (2 hours by default), and so is one whose Run is cancelled. Runs started from the CLI or /v1 have nobody to ask unless they ask for the interactive mode, so there ask means deny. See Permissions.

The audit log​

Every tool call that passes its schema is recorded: the tool name, its arguments and result, approval status and timestamp. Long values are cut short, and values under secret-looking names are replaced by [REDACTED]. The record lives in the encrypted SQLite store (~/.wisp/data.db), so even the audit trail is protected at rest. You can see what the agent did, when, and whether a human approved it. Deleting a Session deletes its audit rows too.

Under the hood​

Local tools

  • Files, in the project folder: reading, listing and searching run without confirmation; writing and editing are ask by default, and so is any call outside the folder while the sandbox is off
  • Shell commands: ask by default
  • Encrypted memory and preferences: store, recall, update and delete
  • Sub-agents: for parallel work, with every tool unless the agent narrows them to reading and bash_run, with no step or time limit unless subagentTimeoutMs sets one (time spent waiting on a confirmation does not count)
  • Skills: invoked by name, with their own files
  • Call recordings: list, read and search

Sandbox. Shell commands can be sandboxed, off by default: with sandbox-exec on macOS and bubblewrap on Linux. A Windows sandbox is in development. See Sandbox.

Remote tools (through the Wisp Proxy, inside the TEE)

  • Web search, page fetching and raw HTTP requests, all executed inside the hardware enclave, results returned over MCP

Connectors. Gmail, Google Calendar, Google Drive, Telegram and any MCP server you add. Their tools are named mcp__<connector>__<tool> and are ask by default. The agent finds them with mcp_search and runs them with mcp_call instead of keeping them all in its context; see Connector tools. See Permissions.

Sessions and memory

  • Chat sessions: encrypted SQLite (~/.wisp/data.db), scoped per project
  • Agent memory: persistent across sessions and Projects, tagged (user, project, feedback, reference), saved by the agent or after a turn on its own, and recalled into each turn by local embeddings or full-text search. See Tools
  • Audit log: every tool call recorded with shortened, redacted arguments and result, timestamp and approval status

Streaming and UI

  • Results stream to the desktop app over Server-Sent Events (SSE)
  • The desktop app renders tool execution and diffs live