What tools do
A tool is something an agent can use to affect the real world. Without tools, an agent can only talk. With tools, it can do.
Examples of tools #
- Reading and writing files on your computer
- Searching the web and reading pages
- Running commands
- Talking to other programs you use
- Controlling your screen, keyboard, and mouse
How an agent uses one #
When an agent decides it needs a tool, it says so, uses it, and gets a result back. All of that shows up in the conversation. You can see:
- Which tool it used
- What it asked for
- What came back
Nothing is hidden. If an agent read a file, you will see which file.
You decide what is allowed #
An agent can only use tools you have allowed. Some are safe and always on. Others ask you first. Some you may want to switch off entirely.
C5 stops and asks before anything risky — deleting things, spending money, sending data off your computer. You get a clear message describing exactly what it wants to do.
This is covered fully in Permissions.
Built-in safety #
Beyond asking you, C5 has guards that work on their own:
- A tool that keeps failing gets switched off for a while, instead of retrying forever. This stops one broken thing from burning your budget.
- Information leaving your computer is checked as a backstop against an accidental leak. It catches common patterns, not everything — do not rely on it to notice a secret.
- Everything is recorded, so you can look back at what happened.
- The same web request is not made twice in one task. While a task runs, a
web_searchorread_urlthat asks for exactly what was already fetched in the last 15 minutes gets the earlier answer, and two steps asking for the same thing at once share one request. Aread_urlthat attaches the page's image always fetches afresh. - A busy search provider is waited for. When Brave search answers that too many requests were made, C5 pauses — for as long as Brave asks, when it says — and tries again, up to three times, holding back other searches meanwhile, instead of failing the step at once.
- A placeholder is never saved as a result.
write_file(and its other name,create_file) refuses to write an empty file, and refuses any text written to a video, image, PDF, Office or ZIP file (.mp4,.mov,.webm,.mkv,.avi,.png,.jpg,.jpeg,.webp,.gif,.pdf,.docx,.xlsx,.pptx,.zip) — except a PDF whose text really is a PDF. Its refusal tells the agent which tool makes the real thing, or to write Markdown that C5 turns into the deliverable. So an agent that cannot make a file says so, rather than leaving a file that only looks like one.
Pictures and video #
Agents can make images and short videos on this machine, and fetch images from the web:
render_svgdraws an image (PNG by default, or JPEG or WebP), from the agent's own drawing or from a ready-made card:editorial_card(a headline card),stat_callout(a big number),quote_card(a quotation) orbadge_banner(a slim header banner, 1200×240 unless asked otherwise, with an optional badge, a one-line title and subtitle, and the source and date).render_videomakes a video from slides: an H.264.mp4, 1920×1080 unless asked otherwise, with each slide's caption shown on it and, for a task, written beside it as a.vttcaption file. It uses ffmpeg on this machine, which C5 does not ship. C5 looks for it first at the FFmpeg location in Settings › Tools, then inFFMPEG_PATH, then on yourPATH, then in/opt/homebrew/bin,/usr/local/bin,/opt/local/binand/usr/bin. Set the location when ffmpeg is anywhere else, or when C5 starts at sign-in, which does not see your shell'sPATHor anFFMPEG_PATHyou exported. It must be the full path to a file namedffmpeg, and only an administrator can set it; a save with anything else is refused and says why. C5 notices a newly installed ffmpeg within a minute, without a restart, and one you removed the next time it looks: it then uses the next ffmpeg on the list, or reports none. Without it, the tool refuses and suggestsrender_svginstead, and a task that asks for a video is told when it is planned that ffmpeg was not found, so it delivers the frames and a storyboard rather than a video.fetch_web_imagedownloads one image from an address into the task, as an attachment with a readable file name and a line naming where it came from. To find a page's main image,read_urlreads the image the page itself names for sharing (its Open Graph, Twitter Card or Schema.org image) and can attach it. Reading a page asks nothing in an unattended run, but aread_urlthat attaches the image saves a file, so it is treated likefetch_web_image: it asks for approval unless a grant or its master switch coversread_url, and the guardian checks it.
A result is kept with the task, up to 25 MB; a larger one is refused, as a video or an image, by its kind. Each of these tools has a switch under Master Tool Authorizations (see Permissions).
Adding more tools #
Two ways:
- Integrations connect C5 to services you already use. See Integrations.
- MCP servers are a standard way to add new abilities. See MCP servers.
Seeing which tools get used #
The Tools view in Analytics shows which tools your agents reach for most, and which ones fail. A tool that fails often is usually misconfigured rather than broken.
Where to go next #
- Permissions — set the rules.
- MCP servers — add new abilities.
- Computer use — let an agent drive your screen.
