simple-agent

simple-agent

Boot a sandbox from a resident app and let an LLM agent use it. This example boots apps/filesystem, turns its exports (write_file, read_file and the rest) into tools, and asks the agent to write a file and read it back. It is the agent-side counterpart to the hello-world resident app.

It uses the experimental agent framework, @berthos/agents, which isn't published. Run it from a clone.

Run it

Needs Docker, and ANTHROPIC_API_KEY or OPENAI_API_KEY. Without a key the script prints SKIP and exits cleanly.

pnpm install && pnpm build            # once, from the repo root
cd examples/agents/simple-agent
export ANTHROPIC_API_KEY=sk-ant-...   # or OPENAI_API_KEY
pnpm start

On a machine whose kernel can't enforce (Docker Desktop on macOS or Windows), the boot is refused. Prefix the command with BERTH_ALLOW_UNENFORCED=1 to run it unenforced, or see enforcement by platform.

The script prints what the agent said and the tools it called.

Two scripts

• index.mjs is the shortest form: runAgent({ apps, task }) with no llm passed. It picks up whichever of ANTHROPIC_API_KEY or OPENAI_API_KEY is set, boots, runs one task and cleans up.
• index-manual.mjs is the fuller form: an explicit LLMProvider, and Agent and Computer handles you keep for more than one turn. It also accepts --connect=<name> to attach to a running berth os up instance.

Skip the boot on every run

Building the image and starting the container on every run adds seconds. Boot once with berth os up, then reconnect:

# from the repo root
berth os up my-agent --apps=apps/filesystem

cd examples/agents/simple-agent
node index-manual.mjs --connect=my-agent   # reconnects in milliseconds

berth os down my-agent                     # from the repo root, when you're done

See the Berth OS reference.

Multi-agent crews

For Crew.withManager() and Crew.networked(), see experimental/agents/examples and the agents reference.