The Pulse
Firecrawl kit gives Docker sandbox agents controlled web access
The Firecrawl kit lets coding agents in Docker Sandboxes search, scrape and crawl the web through a single permitted API host. The kit is experimental and in Early Access, and its access remains subject to sandbox and organization policies.

Coding agents running in Docker Sandboxes can search, scrape and crawl the web through a Firecrawl kit that adds one API host to their permitted network access. Docker described the kit in a November 16 post, saying it returns pages as Markdown while keeping the agent inside the sandbox's network controls.
The kit is experimental and in Early Access, and its format and related commands may change. It extends an existing sandbox workload rather than creating a new one.
How the access works
A coding agent in Docker Sandboxes runs inside an isolated virtual machine, with controls over which hosts it can reach. The Firecrawl kit gives it access to web content through api.firecrawl.dev rather than requiring each website to be added to the sandbox's allowlist.
The kit is a mixin: a component added to a workload. Its installation step adds the Firecrawl Python SDK, firecrawl-py, pinned to version 4.44.0, as the agent user. Sandbox creation fails if the installation or import check fails.
The kit also supplies instructions telling the agent how to use the capability and what limits apply. A proxy supplies the real Firecrawl API key when requests are sent to the API host; inside the sandbox, the corresponding environment variable holds a placeholder instead of the key.
Policies and setup
The network permissions differ by phase. For installation, the kit permits access to pypi.org and files.pythonhosted.org; at runtime, it permits access to api.firecrawl.dev. The kit adds permissions to the sandbox's allowlist but does not remove existing ones.
Docker's balanced preset permits package registries, GitHub and a curated list of vendor domains, while blocking general websites, including most documentation sites. Under a deny-all policy, Firecrawl is the agent's only web host.
The walkthrough requires the sbx command-line tool, version 0.46 or later, signed in with sbx login. It also requires a global network policy, a Firecrawl API key and a workload to which the kit can be added; Docker Desktop is not required. The published kit supports linux/amd64 and linux/arm64.
On the first run, sbx asks for approval to send the Firecrawl credential to the API host. A detached or CI run cannot answer that prompt, so the approval must be given interactively beforehand or the binding written to the credentials configuration.
What agents can retrieve
In Docker's example, an agent is asked to read a Pydantic changelog and list the latest release's changes. The kit's instructions direct the agent to scrape blocked hosts through Firecrawl, and the example shows it fetching the page rather than answering from memory.
The same access can be used to search for unfamiliar errors, read API references or crawl sections of documentation.
Limits remain
Under the balanced preset, general sites still return a proxy error when accessed directly, so a link found in a scrape result must also be retrieved through Firecrawl. Screenshots and other media return links on a Firecrawl storage host that is not on the allowlist, leaving those files unavailable for download inside the sandbox.
An organization's managed policy can override the kit's allow rule. In that case, scrapes fail until the organization permits api.firecrawl.dev. An HTTP 402 response means the Firecrawl account has run out of credits, and the agent is instructed to stop rather than retry.
Source
Related stories
- Entire gives AI agents service accounts for repository access
- LangChain puts agent records at the center of LangSmith
- Vercel opens paid Security Review to unlinked GitHub repositories
- H Company makes AI agents available on AWS European Sovereign Cloud
- Cloudflare open-sources experimental Kitesurf browser for AI agents