Skip to content

Tools ​

A tool lets an agent do something besides answer. On an agent's Configure tab, they are switches under Abilities:

Every agent can search the web and read web pages, wherever its space has them on, from the day it's created. An editor can turn either off for one agent, at once and without publishing: for one that works with things that shouldn't leave, for instance.

Computer is off until three separate people have said yes: the platform offers it and sets ceilings on it, the owner (you, or your organization's admins) turns it on and configures it, and an editor gives it to a particular agent and publishes. That is deliberate: running code is a different kind of power.

All three are in the space's Settings, under Web and Computer, with their settings folded under each. There is nothing to add: they are there in every space.

Every call is shown in the chat as a card and recorded in the owner's audit log, with the command or URL and whether it succeeded.

Asks a search engine a question and gives the agent the best results: title, address and a short extract of each. Pair it with web fetch and an agent can find a page and then read it, instead of being told which page to read.

On a server that runs its own search engine (crewhall.space does, and so does a standard install), it's on in every space, free and with nothing to set up: This server's search asks the usual search engines for you. A space can switch it off, or choose another engine in its card's settings:

EngineNeedsGood when
This server's searchnothingthe default, when the server runs one
Brave Searchan API key from their developer siteyou want a hosted engine with a free tier
Tavilyan API keyyou want results already trimmed for agents
SearXNGthe address of one you runyou would rather no query left your network

The key is encrypted here, never shown again and never given to a model. Changing engine clears it, since a key from one is no use to another. Until there is something to search with, the tool is not offered to agents at all, rather than failing when one tries.

Results are things found, not instructions

Titles and extracts are written by whoever made the page, and a page can be written to talk to your agent. Agents are told to treat results as things found. The agent still cannot open any of them unless web fetch is on (it is by default) and the host isn't blocked.

Web fetch ​

Reads a web page or a JSON endpoint, and gives the agent the text.

It's on in every space where the platform offers it, and reaches any public website: an agent can research a question, follow a link from a search result, read the docs it's pointed at. A space's owners and admins can narrow it in its card's settings, for what a project needs:

  • Allowlist: when it has hosts, only those are reachable.
  • Blocklist: these are never reachable, allowlisted or not.

Both also apply to redirects, and to what computers reach. Patterns are exact hostnames or *.example.com for subdomains.

When an agent asks for several pages or searches in one go, they're fetched at the same time (up to 8 at once), so a brief that reads ten pages waits for the slowest one, not for all ten in turn. Telling an agent to "fetch all of these at once" in a task's instruction helps it ask that way. Owners can also switch the tool off.

Private addresses (localhost, 10.x, 192.168.x) are never reached by "any website": only a host on the allowlist may resolve to one, and only if the platform allows it and the owner turns it on, which only makes sense on a self-hosted install where every owner is trusted. Cloud metadata addresses stay blocked regardless.

What comes back is not an instruction

A page can contain text addressed to the agent, trying to get it to do something. Agents are told to treat fetched content as data. If an agent works with things not everyone may see, an allowlist keeps it to the sites its work needs.

Computer ​

Gives the agent a small Linux container of its own (the shell tool), with a /workspace directory whose files survive between commands and between conversations. Debian, with Python, git, curl, jq and ripgrep.

Good for: writing and running a script, working through a file, keeping notes or intermediate results across a long job.

What bounds it:

  • Its own network: none. The container can't see the server it runs on, the other agents' computers or anything else. Its one way out is the filter below.
  • Memory, CPU and a per-command time limit, set by the owner within the platform's ceilings.
  • An ordinary user, no privileges, and only its own volume to write to.

The computer belongs to the agent, not to you. Everyone who talks to that agent shares those files, in the same way they share its memory. Deleting the agent deletes them.

The internet, filtered ​

A computer reaches websites through CrewHall's filter, a proxy every request goes through:

  • Websites only: http and https, on ports 80 and 443. No ssh, no mail, no other ports, not even DNS: the filter looks names up itself.
  • Public addresses only, as for web fetch: never the server's own network, a private address or a cloud metadata address. A site whose name points to one is refused.
  • The space's lists: web fetch's allowlist and blocklist apply here too. With an allowlist, an agent reaches only those sites, so a computer meant for one job can be kept to it (say pypi.org and *.pythonhosted.org for Python packages).

curl, pip, git (over https) and most tools use the filter by themselves; the agent is told how it works. It is on wherever Computer is, and follows the agent's Read web pages switch: off, its computer is offline too. A space's owners can take it from every computer at once in Computer's settings (Internet for computers), and the platform can take it from every space.

What a computer with the internet can do

It can download code and run it, and send what it reads to any site it may reach, in a request it builds itself. Agents read web pages and documents written by others, and a page can be written to talk them into things. If an agent works with things that shouldn't leave, turn its Read web pages off, or keep its space to an allowlist of the sites its work needs.

Turning tools on ​

  • Settings → Web and Settings → Computer: what your agents may use, and how it is configured. Switching web search or web fetch off here takes it from every agent in the space.
  • Agent → Configure → Abilities: Search the web and Read web pages for this agent (at once), and Computer (then publish).

The space's Settings: web search and web fetch under Web, and Computer with its limits and internet

If a tool is switched off at the owner level, agents stop being offered it at once, and an agent that has Computer can no longer be published until it is removed.

CrewHall