# Tool inventory: 75 built-in tools and your permission boundary

> NearWork ships 75 built-in tools: file editing, command execution, web access, documents, images, audio, browser and desktop control, SSH, memory. Plus the four permission modes.

"What can the AI actually do" is really a question about which tools it can call. NearWork has **75 built-in tools**, covering everything from reading a single file to driving a desktop application.

## Tools are loaded on demand, not dumped up front

Filling the context with every tool on a cold start is expensive and makes the model choose badly among dozens of options. NearWork loads a small core set and lets the model call `enable_tools` to pull in what the current task needs before calling the specific tool. You can also name the capability in your prompt: "open this URL in the browser and screenshot it".

## Grouped by capability

**Files and code**: `file_read`, `file_write`, `line_edit`, `insert_lines`, `delete_lines`, `replace_in_file`, `batch_file_edits`, plus `glob`, `grep`, `list_dir` for search and `diff_preview` before changes land. Editing by line rather than rewriting whole files is why long-file edits stay reliable.

**Commands and processes**: `run_command` for one-shot jobs; long-running work goes through `process_spawn` and is inspected or stopped with `process_read`, `process_send`, `process_kill`; `sleep` for waiting.

**Web and network**: `web_search`, `web_fetch` for page text, `http_request` for APIs, and `lingxi_open_web` to show a page in the side panel while you work.

**Documents and spreadsheets**: read and write for CSV, Excel, Word and PDF (`read_csv`, `write_xlsx`, `read_docx`, `write_pdf` and friends), with `ocr` for images and scans.

**Images and audio**: `image_understand`, `image_generate`, `image_edit`, `image_process` for resize and crop, `audio_tts` for speech synthesis and `audio_stt` for transcription.

**Desktop and browser control**: `lingxi_computer` drives a GUI through screenshot-and-vision loops (clicking, typing, filling forms across applications), `lingxi_mouse` and `lingxi_keyboard` perform precise single steps, `lingxi_browser` opens the built-in browser, `take_screenshot` captures the screen.

**Remote and developer**: `ssh_exec` and `ssh_sftp` for remote hosts, `git_status` and `git_diff` for repositories, `subagent_spawn` to fan work out in parallel.

**Tasks, collaboration, misc**: `task_create` and `task_schedule` for to-dos and recurring jobs, `ask_user` to stop and confirm with you, `compress` / `extract` for archives, `clipboard_read` / `clipboard_write`, `notify` for system notifications.

**Memory and knowledge base**: a set of `memory_*` and `knowledge_*` tools that keep long-term preferences and local documents around, retrieved through the embedding slot.

## Four permission modes

The stronger the tool, the clearer the boundary needs to be. For writes, commands and desktop control, NearWork asks first:

| Mode | Behaviour | When to use |
|---|---|---|
| `normal` (default) | Read-only automatic; writes and commands confirmed | Everyday use |
| `strict` | Confirmed at every step | Unfamiliar projects, production directories |
| `auto` | Fully automatic; dangerous commands still blocked | Reversible, reviewable work |
| `full_access` | Nothing is confirmed | Disposable sandbox only |

Slower is better than sorry: do not run `auto` or above against directories that matter. Even a good model can misread your intent.

## Extending beyond the built-ins

MCP servers add third-party capabilities (browser automation, documentation knowledge, collaboration tools, databases), and custom skills package a repeatable procedure into one sentence. Configuration is covered in the [MCP guide](/en/docs/features/mcp).

## Frequently asked questions

### Are more tools always better?

No. The right number is "just enough": unrelated tools dilute attention and invite mis-calls. What actually decides the outcome is whether your prompt states the goal and the constraints.

### Do tool calls cost extra?

Tools themselves are not billed separately, but the tokens a model spends do count toward usage, and web search and fetch are metered per request through the relay. See the [pricing page](/en/pricing).

### Can I disable specific tools?

Yes. Permission modes block writes and commands, and removing a server from the MCP config removes its tools entirely; tools that were never enabled never enter the context.

## Next steps

- Want to know which models sit behind these tools? See [Models and providers](/en/docs/models/providers).
- Care where data crosses the network? Read [Data and privacy](/en/docs/faq/data-privacy).
- Not installed yet? Go back to the [quick start](/en/docs/start/quick-start).
