agent-network-access-probe
Supports Docker as a sandbox runner, executing scanned agent tools in an isolated Docker container to collect runtime network-egress evidence and findings without requiring OpenShell.
Supports local Ollama models as an OpenAI-compatible LLM endpoint for generating AI fixes without relying on a cloud model provider.
Uses OpenAI-compatible LLM endpoints, including OpenAI models, to generate AI fixes for static-analysis findings while still producing policy, profile, findings, and attestation output without a model.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
Probe
Public product: Trustabl Probe
Runs your agent tool in an isolated OpenShell sandbox, observes the network destinations it actually reaches, and generates a least-privilege OpenShell network_policies draft from that evidence — no hand-written allow list required. Static analysis also flags hardcoded credentials, missing timeouts, and other reliability issues, with AI-generated fixes for each, reviewed one by one. Your source code never leaves your environment.
Language support: Static analysis is supported for Python, Go, JavaScript/TypeScript, Ruby, and Shell. Sandbox execution is available for all languages.
How It Works
Your tool is uploaded to a sandboxed OpenShell environment and executed
Runtime evidence is collected — network destinations reached, denials, crashes, warnings, policy violations
Static analysis scans your source for hardcoded credentials, unauthorized network calls, missing timeouts, and more
You get back a least-privilege OpenShell policy YAML (unsigned draft) built from what the tool actually reached, findings from both passes, and a fixed version of the code for each issue found
Each session can also be attested: an optional signed manifest of the run, using Sigstore keyless signing — no private key to manage. In CI (e.g. GitHub Actions with permissions: id-token: write), it signs with your workflow's identity automatically. Run locally without that, and it falls back to an interactive browser login bound to your personal identity, recorded in Sigstore's public transparency log — pass --no-sign to skip signing entirely instead.
Cloud metadata endpoints and link-local addresses (169.254.169.254, metadata.google.internal, 169.254.0.0/16, fe80::/10) are always excluded from the generated policy, regardless of what the tool actually reaches — pass --deny-list FILE to add your own hosts/IPs/CIDRs on top of that built-in set.
By default, everything a tool observably reaches (and isn't denied) goes into the draft. Pass --allow-list FILE to make that stricter: once supplied, an observed destination not on the list is flagged instead of auto-added — the allow list becomes the required source of truth, not just an addition to what was seen. --platform-allow-list FILE works the same way for platform must-reach endpoints (e.g. NIM, IdP) and tags matching entries source=platform in probe.profile.json.
If a tool's code contains a call shape with a real-world side effect — a network write, a subprocess call, a filesystem delete, destructive SQL — its sandbox execution is skipped by default (static analysis still runs and still reports on it) so a default run can never actually charge, delete, or modify anything. Pass --allow-write-tools to run it for real once you're ready, ideally alongside --env-fixtures FILE so the tool picks up staging/test-mode credentials instead of production ones.
Sandbox results are cached per tool under os.UserCacheDir()/trustabl-probe/results — keyed by the tool's source content and effective run config (base image, --sandbox-args, --env-fixtures content), not its filesystem path, so the same tool referenced by two different agent classes reuses one cache entry. probe.profile.json's cache_hit field reflects whether each tool's result came from the cache. Pass --no-cache to skip the cache entirely (neither read nor write) — always do this for release gates.
Related MCP server: Trustabl for Cursor
Requirements
A sandbox runner for
--sandboxexecution — pick one with--runner:docker(default-friendly, no separate install beyond Docker itself) — Docker Desktop or Docker Engine, daemon running. No OpenShell required.openshell(default--runner) — OpenShell installed, withopenshell gateway startrunning. Also requires Python 3.8+ on the host machine, for dependency wheel downloads.
Access to an OpenAI-compatible LLM endpoint (OpenAI, Anthropic, Ollama, LiteLLM, LM Studio, etc.), only for generated fixes. Without
--model, Probe still writes the policy, profile, findings and attestation.
Setup
1. Install a sandbox runner
Docker only (no OpenShell needed): install Docker Desktop or Docker Engine and make sure the daemon is running (docker info). Pass --runner docker to fix/watch — that's the whole setup.
OpenShell:
Download from build.nvidia.com/openshell, then:
openshell gateway start # keep running in a separate terminal
openshell sandbox list # verify CLI is on PATH2. Configure an LLM (optional, only for generated fixes)
Local (Ollama):
brew install ollama # macOS
ollama serve
ollama pull qwen2.5-coder:32bCloud: set OPENAI_API_KEY or ANTHROPIC_API_KEY — the endpoint is auto-detected from the model name.
3. Install autofix
macOS (Homebrew):
brew tap trustabl/autofix
brew install autofixWindows (Scoop):
scoop bucket add trustabl https://github.com/trustabl/scoop-autofix
scoop install autofixLinux / manual:
# Replace v0.1.0 with the latest version from https://github.com/trustabl/agent-network-access-probe/releases
curl -L https://github.com/trustabl/agent-network-access-probe/releases/download/v0.1.0/autofix_0.1.0_linux_amd64.tar.gz | tar xz
sudo mv autofix /usr/local/bin/
autofix versionOr use it from your coding agent (MCP)
Probe also runs as an MCP server, so a coding agent (Cursor, Claude Code, Claude Desktop) can check what Probe needs, install it, and scan a tool, then apply the fix hints itself. No model key is needed. See mcp/README.md.
Commands
fix — Generate a least-privilege policy from observed egress, and fix what's found
autofix fix --source ./my-tool
autofix fix --source ./my-tool --entrypoint tool.py --sandbox --model gpt-4o
autofix fix --source ./my-tool --existing-policy ./policy.yaml --model claude-3-5-sonnet-20241022Flags:
Flag | Default | Description |
| required | Path to the tool directory |
| auto-detected | Entry point script(s) — can repeat for multi-file scanning |
| — | Model name. Cloud: |
| auto-detected | OpenAI-compatible base URL. Auto-detected for |
| env var | API key — defaults to |
|
| Per-request LLM timeout (increase for slow local models) |
| off | Run the tool in a sandbox for runtime findings |
|
| Sandbox runner: |
| auto | Sandbox base image (default: auto-detected from entrypoint extension) |
| — | Arguments passed to the tool inside the sandbox. Value is a raw shell token — avoid shell metacharacters. |
| — | Path to an existing OpenShell policy YAML to diff against the generated least-privilege policy |
| — | Path to a deny list file (one host/IP/CIDR per line, |
| — | Path to a declared allow list file (one host/CIDR/ |
| — | Path to a platform allow list file (jail must-reach endpoints, e.g. NIM/IdP) — same format as |
| off | Allow sandbox execution of tools that appear to perform mutating/irreversible operations (default: skipped, static analysis only) |
| — | Path to a dotenv-style file ( |
| off | Skip the per-tool result cache entirely — neither read nor write it (use for release gates that must always execute for real) |
| off | Accept all fixes without interactive review (useful for CI) |
| off | Skip Sigstore signing — no network call, no interactive browser flow, no public transparency log entry |
|
| Output format: |
Output files (written to --source directory):
File | Description |
| Least-privilege OpenShell network policy (unsigned draft, built from observed egress) |
| Per-tool egress table, grant-set hash, and scan metadata (unsigned) |
| Every finding (title, severity, evidence, fix hint) with file attribution — always written, independent of whether interactive fix review ran |
| Optional signed behavioral manifest with source hash |
| Full session log (findings, diffs, accept/reject decisions) |
| Unified diff of all accepted fixes |
watch — Continuous improvement loop
Polls source files for changes. On every save, re-runs static analysis, rewrites network_policies.draft.yaml, probe.profile.json, probe.findings.json, and autofix-attestation.json, and prints a delta (score change, resolved/new findings, new source hash).
autofix watch --source ./my-tool
autofix watch --source ./my-tool --entrypoint tool.py --interval 1sFlags:
Flag | Default | Description |
| required | Path to the tool directory |
| auto-detected | Entry point script(s) to watch |
|
| Poll interval |
| off | Skip Sigstore signing — no network call, no interactive browser flow, no public transparency log entry |
| — | Path to a deny list file (one host/IP/CIDR per line, |
| — | Path to a declared allow list file (one host/CIDR/ |
| — | Path to a platform allow list file (jail must-reach endpoints, e.g. NIM/IdP) — same format as |
verify — Verify an attestation and gate CI
Checks an attestation's content digest and Sigstore signature, then optionally fails (non-zero exit code) on a low score or critical findings. See docs/ci-gate.md for a full CI setup.
autofix verify ./autofix-attestation.json --min-score 70 --fail-on-critical
autofix verify ./autofix-attestation.json --require-signature --identity "https://github.com/acme-corp/.*"Flags:
Flag | Default | Description |
|
| Fail if |
| off | Fail if any critical findings are present |
| off | Fail if no valid Sigstore signature is present |
| — | Regex the signer's certificate SAN must match |
| — | Regex the signer's OIDC issuer must match |
|
| Output format: |
version — Print version
autofix versionlicense — Print license and copyright information
autofix licenseLicense
Apache License 2.0. See LICENSE, or run autofix license to
read it in the terminal.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Runtime behavioral trust scoring for MCP servers. Check reliability before calling unknown tools.
Guarded MCP server for agent-readable business truth, provenance, readiness, and discovery.
Secure Linux sandboxes for AI agents, as MCP tools.
Scan any MCP server for tool-poisoning, security, auth & license. Trust score before install.
Related MCP Servers
- AlicenseAqualityDmaintenanceMCP server for verifying high-impact decisions with Trust OS.2MIT

Trustabl for Cursorofficial
FlicenseNot gradedqualityCmaintenanceMCP server that brings Trustabl static reliability scanning into Cursor, letting agents inventory agents/tools, flag risky patterns, score production readiness, and write SARIF reports.30-- AlicenseNot gradedqualityCmaintenanceEnables security auditing of MCP servers by running them in a sandbox with fake secrets, capturing outbound traffic, and detecting tool poisoning or secret exfiltration before approval.MIT
- AlicenseCqualityBmaintenanceProduction-grade MCP server providing deterministic evaluation, AST sandboxed execution, DNS routing audit, and runtime guardrails for autonomous LLM agents.1MIT