About Trustabl Probe
Trustabl Probe is an MCP server that statically inspects agent tool code and runs tools in a network-blocked sandbox to find runtime issues. It drafts a network allow list for coding agents and returns findings with fix hints.
Key features
- 01Check Probe and sandbox readiness
- 02Download and verify the Probe binary
- 03Inspect code and draft a network allow list
- 04Run tools in a sandbox to find runtime issues
- 05List findings by severity with fix hints
- 06Write scan reports to the target directory
Use cases
- 01Review network access needs for agent tools
- 02Find crashes and missing dependencies in a restricted network
- 03Check tool code for credentials or missing timeouts
MCP support
The MCP server exposes setup-check, installation, and scanning tools to clients, including probe_check_setup, probe_install, and probe_scan.
Configuration
Local server config is the same for every client. Copy it into your client’s MCP config file.
We haven’t collected a package or remote endpoint for this server. Follow the instructions in the project README.
Tools
3Saved tool list · updated 1d ago
This tool takes no parameters.
Details
From the project READMETrustabl Probe MCP server
Gives your coding agent Trustabl Probe: it drafts a network allow list for an agent tool and runs the tool in a network-blocked sandbox to catch problems that only show up at runtime. Every finding comes with a fix hint, and your agent applies the fixes, so no model or API key is needed.
Tools
| Tool | What it does | Changes anything? |
|---|---|---|
probe_check_setup | Reports whether the Probe binary, Docker and the sandbox images are ready, which languages can run in the sandbox, and the exact command to fix each missing piece. | No |
probe_install | Downloads the Probe binary (checksum-verified) and pulls the sandbox images. For Docker, OpenShell or Python it only returns the command to run, with its caveats. | Yes, see below |
probe_scan | Scans a tool directory and returns a readiness score, the draft network allow list, and findings worst-first with fix hints. | Writes four report files into the scanned directory |
Static and sandbox
- Static reads the code and flags problems such as missing timeouts, hardcoded credentials or unfiltered tool output. It also drafts the network allow list (an OpenShell network policy) from the code. It always runs and needs nothing extra.
- Sandbox also runs the tool with its network blocked, catching crashes,
missing dependencies and blocked connection attempts. It needs Docker (or
NVIDIA OpenShell). With
sandbox: "auto", the default, it runs whenever a runner is ready. Tools that look destructive (payments, deletes, network writes) are not run unless you passallow_write_tools: true.
Docker covers Python, JavaScript/TypeScript, Go and Ruby; the OpenShell image runs Python and Node. Shell scripts are always analyzed statically only.
Install
Requires Node.js 18 or later. Docker is optional and enables the sandbox.
The server ships with every Probe release
as trustabl-probe-mcp-<version>.tgz. The examples use v0.2.4; use the latest
release.
Cursor (.cursor/mcp.json):
{
"mcpServers": {
"trustabl-probe": {
"command": "npx",
"args": ["-y", "https://github.com/trustabl/agent-network-access-probe/releases/download/v0.2.4/trustabl-probe-mcp-0.2.4.tgz"]
}
}
}
Claude Code:
claude mcp add trustabl-probe -- npx -y https://github.com/trustabl/agent-network-access-probe/releases/download/v0.2.4/trustabl-probe-mcp-0.2.4.tgz
Claude Desktop: download trustabl-probe-mcp-<version>.mcpb from the release
and open it to install. Or use the same command and args as the Cursor example
under mcpServers in claude_desktop_config.json.
Then ask your agent to run probe_check_setup.
What it changes on your machine
- Downloads the Probe binary from the same release as the server into
~/.trustabl-probe/<version>/, after checking it against the releasechecksums.txt. A mismatch is a hard failure. - Pulls the sandbox images (
python:3-slim,node:20-slim,golang:1.25-alpine,ruby:3-slim) when you runprobe_install, or on the first sandbox run. - Writes
network_policies.draft.yaml,probe.findings.json,probe.profile.jsonandautofix-attestation.jsoninto the scanned directory. - Never installs Docker, OpenShell or Python itself. It returns the command, and your agent should ask you before running it. Docker Desktop needs admin rights, may need a restart on Windows, and needs a paid subscription for larger companies.
Privacy Policy
Trustabl's privacy policy: https://trustabl.ai/privacy/
What this server does with your data:
- Collection: none. It sends no telemetry and makes no model call.
- Your code: it is read and analyzed on your machine and never uploaded. Attestation signing is turned off, so nothing goes to a transparency log.
- Network access: only GitHub (to download the Probe release), Docker Hub (for the sandbox images) and, when a Python tool declares dependencies, PyPI to install them before the sandbox run. The tool under test runs with its network blocked.
- Storage: the Probe binary is cached in
~/.trustabl-probe/, and the four report files are written into the scanned directory. Delete either at any time. - Third parties: nothing is shared with Trustabl or anyone else.
- Contact: https://github.com/trustabl/agent-network-access-probe/issues
Settings
| Variable | Default | Purpose |
|---|---|---|
PROBE_VERSION | the release the server came with | Probe release to use instead, e.g. v0.2.3 or latest. Needs v0.2.3 or later. |
PROBE_BIN | Path to a local Probe build; skips the download | |
GITHUB_TOKEN | Only to avoid GitHub's anonymous rate limit when PROBE_VERSION is latest |
Development
From mcp/:
cd server && npm ci && npm run build # bundles to dist/index.js, which is committed
node test/smoke.mjs # handshake, tools/list, setup check
SMOKE_SCAN=1 node test/smoke.mjs # also a static scan of test/fixture with the released Probe
PROBE_BIN=/path/to/autofix node test/smoke.mjs # the same scan with a local Probe build
A checkout has version 0.0.0 and uses the latest Probe release. The release
workflow sets the real version when it packs the server.
License
Apache-2.0, like Probe.
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.