Analyze a suspicious file with an agent

Give an agent a suspicious file, have it inspect the bytes inside a temporary Capsem VM, and retrieve a report on your computer. This walkthrough uses a harmless text fixture containing suspicious-looking strings. It performs static triage: it does not execute the sample or contact its URLs.

Install the MCP server

Install Capsem and start it on your computer. With Node.js 20 or later installed, run:

npm install --global @capsem/mcp@0.7.0

This installs capsem-mcp, which gives your agent tools to create VMs, run commands, and transfer files. The package was tested against Capsem 0.7.1.

Connect your agent to Capsem

In your agent’s MCP settings, register a stdio server with these fields:

Field	Value
Executable	capsem-mcp
Arguments	--gateway-url, http://127.0.0.1:19222, --token-file, /absolute/path/to/.capsem/run/gateway.token

Replace the token-file path with the absolute path under your home directory. If Capsem uses another port, read it from ~/.capsem/run/gateway.port and replace 19222. The server reads the token file on the host; do not paste its contents into the agent’s prompt or upload it to the VM.

Reconnect the agent and check that it lists capsem_create, capsem_exec, capsem_write_file, capsem_read_file, and capsem_delete. If they are missing, check MCP startup diagnostics before trying the task. Exact schemas are in the MCP reference.

Install the agent skill

In your agent’s project directory, run:

npx skills add google/capsem --skill use-capsem

Select your agent when prompted. The use-capsem skill teaches VM creation, file transfer, bounded execution, result checks, and cleanup. It includes the training sample and the Python inspection helper used below. The skills CLI also supports selecting an agent with --agent.

Ask for a static report

Save the harmless sample and inspection helper into your project directory. Give your agent this task:

Use the use-capsem skill and Capsem MCP to inspect triage-sample.txt.
Create a new temporary VM without an image. Upload the sample and triage.py
using capsem_write_file. Run only:
python3 /root/triage.py /root/triage-sample.txt /root/report.json

Do not execute the sample or follow its URLs. Do not attach a network,
publish a port, or provide credentials. Check the command's exit code.
Retrieve report.json through capsem_read_file and save it as report.json
in this project. Retrieve all bytes if the result is truncated.
Delete only the temporary VM you created after retrieving the report.
Summarize the observed hash, file size, strings, and URL indicators.
Distinguish observations from conclusions about whether the file is malware.

The helper reads the sample as data using Python’s standard library. It writes the SHA-256 hash, size, a file-signature hint, printable strings, and extracted URLs. For a real PE/ELF sample, use a format-specific static parser to inspect imports and sections; this helper does not claim that depth of analysis.

Check the returned report

Open the local report.json. It should contain the sample’s SHA-256 hash, executed_sample: false, network_requests: false, and this URL indicator:

https://updates.example.invalid/payload

The fixture also contains a PowerShell command string. Finding those strings is an observation; it does not prove execution, malicious intent, or safety. The flags describe this helper’s operations, not a universal record of every process in the VM.

Ask the agent to confirm successful capsem_delete for its returned VM ID. Keep the retrieved report before deleting the VM. To inspect recorded activity before cleanup, use Investigate a VM.