Skip to content
DuckTapeKiller edited this page Sep 13, 2026 · 1 revision

Dive runs a server on your computer, reads web pages and files for the model, and can run code. This page explains what each part can reach and which checks stand between the model and your computer, so you can decide what to switch on.

What Dive assumes

  • You are trusted. It's your computer and your data.
  • The model isn't. Text it reads, such as a web page, a document in your library or a paper it fetched, can steer it. Dive treats the model's tool calls as requests to check, not orders.
  • Web content isn't. Pages are evidence, never instructions. The tool descriptions tell the model so, and the checks below don't depend on the model listening.

The local server

  • Dive's server listens only on 127.0.0.1.
  • It refuses any request whose Host isn't 127.0.0.1 or localhost on Dive's port, or whose Origin points anywhere else.
  • It sends a Content-Security-Policy with a fresh nonce for every page, and forbids other sites from framing it.
  • There's no login: listening only on your own computer is the protection. Never expose Dive's port to a network.

Web access

Every web request the tools make goes through one guard: web_scraper, http_request, deep_research and the other research tools, and the agent browser. The guard:

  • only allows http and https;
  • blocks localhost names and private or local addresses: 0.0.0.0/8, 10.0.0.0/8, 100.64.0.0/10, 127.0.0.0/8, 169.254.0.0/16, 172.16.0.0/12 and 192.168.0.0/16, plus IPv6 loopback, unique local (fc00::/7) and link-local (fe80::/10) addresses and IPv4 addresses written in IPv6 form;
  • looks up the host name and checks every address it resolves to, so a public name that points at an internal address is refused as well.

http_request checks the address again before following each redirect, and the agent browser checks every request a page makes, not only the page you open. A blocked request returns "Access to local or private network addresses is not allowed." deep_research never uses Grokipedia.

Files

  • file_operations only works inside ~/dive/workspace.
  • code_search and git_tools only read, and only inside the workspace and the folders listed in ~/dive/allowed-dirs.json. A path outside them is refused with a message naming the allowed folders.
  • shell_command starts in your home folder or an allowed folder, but a shell command can reach anything, which is why it asks first and starts switched off.

Uploads are capped at 50 MB, with at most 8 images per message. Uploaded images are stored under a hash of their content, never under the name they arrived with.

Code and programs

What Can reach What stands in the way
shell_command Anything your user account can do Off by default. Asks every time
run_code Node.js built-in modules, in a worker thread that isn't a sandbox Asks every time
run_python Anything Python can do Asks every time
macos_control AppleScript, opening apps and files, notifications, processes Off by default. Asks every time
browse_act Clicks and typing on the page the browser has open Off by default. Asks every time
Custom shell tools Anything your user account can do Ask every time
Custom JavaScript tools Node.js built-in modules, in a worker thread Nothing: they don't ask
External tools Everything, because they run inside Dive's server Only ask if they set requiresConfirmation
MCP servers Whatever the program does Started only through npx, node, python, python3, uvx, deno or bun, never with flags such as -e or --eval. Their tools don't ask
Skills Nothing They can't run anything, and activate_skill only reads inside the skill's folder

Plugins written by the model

propose_plugin starts switched off. When it's on, the model can only write a draft into ~/dive/plugin-drafts, which nothing loads. The draft runs only after you approve it in Settings > TOOLS, and that approval is the moment the model's code gets full access to your computer. Read the code first.

Permission prompts

Each run of a risky tool opens a dialog showing the command, the code or the inputs, with ALLOW, DENY and CANCEL. If nobody answers, the timeout settings on Pi mode's MAIN tab apply in every mode: after PERMISSION DECISION TIMEOUT, 45 seconds by default, Dive applies TIMEOUT DEFAULT ACTION. Leave that on Deny. With Allow, an unanswered request runs.

Every request, run, refusal and expiry is written to ~/dive/security-events.jsonl. See How the model uses tools.

Secrets and logs

  • Cloud API keys are stored as plain text in ~/dive/cloud-settings.json, protected only by the file's permissions. The server never sends them to the window. You can use environment variables instead; see Cloud mode.
  • Execution traces saved with conversations are built from chosen fields, length-capped, and scrubbed by redact.js before they're written, because tool output can quote a credential.

Pi mode

pi has its own permissions. If you use the pi-sandbox extension, it reads ~/.pi/agent/sandbox.json and .pi/sandbox.json in the project, and Dive's Pi settings show whether each is on. Dive shows pi's questions in its permission dialog. Dive's own tools, skills and MCP servers never reach pi.

A sensible setup

  • Keep shell_command, macos_control and browse_act off unless you need them.
  • Leave TIMEOUT DEFAULT ACTION on Deny.
  • List only the folders you need in ~/dive/allowed-dirs.json.
  • Read an external tool's code before installing it, and a draft's code before approving it.
  • Only add MCP servers you trust.
  • Never expose Dive's port to a network.

For developers

  • The web guard and folder rules are in skills/sandbox.js: assertUrlAllowed(), urlGuardError(), isBlockedIp(), resolveAllowedPath() and loadAllowedDirs().
  • The tools that ask are GATED_BUILTIN_SKILLS in skills.js, together with requiresConfirmation on external tools and the custom shell type. The check runs in executeToolCallWithConfirmation() in routes/chat.js.
  • The request check is isRequestAllowed() in server.js, and the upload limits are MAX_UPLOAD_PAYLOAD_SIZE and MAX_TURN_IMAGES.
  • Trace scrubbing: redact.js and server/trace.js.
  • docs/security.md goes further into the threat model.

Read next

Clone this wiki locally