Prompts & messages
Structured rules, entropy checks, PII patterns, and custom terms on model-bound inputs.
Skilak Mesh inspects prompts, files, images, and tool-call payloads, then blocks or redacts sensitive data before it leaves your boundary.
Agent output
Mesh inspection
Provider ingress
Audit log
Agent output
Mesh inspection
Provider ingress
Audit log
Agent output
Mesh inspection
Provider ingress
Audit log
Coverage
Sensitive information reaches AI systems through more than chat boxes. Mesh inspects every supported, reroutable request before it reaches an external provider.
Structured rules, entropy checks, PII patterns, and custom terms on model-bound inputs.
Inspection of model-bound JSON, tool arguments, metadata, and agent messages.
Bounded text, metadata, rendered-page OCR, and embedded-content inspection.
Bounded OCR, metadata, raw-string, frame, and QR/barcode inspection.
Recursive scanning of ZIP, TAR, and EML files to enforce policy on nested contents.
Relayed to the client unmodified and not inspected. Mesh is an egress control: it enforces policy on what you send and records response metadata payload-free.
Complements OS sandboxing by inspecting outbound traffic routed through the mesh. It does not monitor filesystem reads.
Compatibility
Search your tool. Mesh inspects outbound requests from any client you can point at a custom API base URL. If a tool cannot be routed, the protection is blocking it at egress and using an approved alternative.
32 tools
No tools match your search.
Check the one rule below for the universal answer.
If your tool lets you set a custom API base URL, point it at the mesh alias path and it is covered. If it is a closed app with no custom-endpoint option, the answer is block at egress and use an approved routed client instead.
The /p/<alias>/ prefix is required: the alias is the only thing that selects an upstream, and a bare http://localhost:51737 base URL is refused with unknown_provider_alias. The client keeps and sends its own provider credential; the mesh relays it to the pinned provider and holds none of its own.
No tool can claim all AI traffic is protected on an unmanaged machine. Supported outbound requests routed through the mesh are inspected; managed egress policy blocks known direct AI paths.
Policy
Define policies hierarchically. Mesh resolves precedence across teams, data categories, and global defaults to compute the one correct action.
Evaluation precedence
Files
Mesh unpacks archives, extracts text from documents, and runs OCR on images to apply policy across formats. Anything it cannot safely process is stopped, not forwarded.
> unzip -l data-export.zip Archive: data-export.zip ├── config.json ├── README.md ├── .env (BLOCKED) └── src/ └── main.py ERR: Policy violation in .env
Architecture
Mesh runs as a standalone gateway. It does not require modifying agent code or sharing policy evaluation with an external service.
The scanner evaluates content independently of the destination. It holds no API keys and makes no outbound calls for policy resolution.
The request is scanned, audited, and enforced before it is forwarded. Provider responses are relayed unmodified and are not inspected.
Audit records hold metadata about violations (policy ID, byte count, latency), never the sensitive payloads. Your security tool should not become a toxic data spill.
Inspection failures and unscannable content block by default. An explicit fail-open setting is limited to documented scanner failures and is recorded as a warning.
Agent safety
Mesh can inspect data when an agent sends it through the gateway. It cannot make unrestricted filesystem access inherently safe. Least privilege comes first.
The agent runs in an isolated container or VM.
Only necessary files are mounted to the workspace.
Environment variables and config files are removed from context.
Network blocked except for the route to Mesh.
Short-lived tokens issued through a secure metadata service.
Final boundary scan of prompts, tools, and files.
Quickstart
One policy-controlled boundary between your AI tools and their providers, running in minutes. Clients only need a custom API base URL.
# Install Skilak Mesh (Homebrew)
brew install skilaklabs/tap/smesh
# Start the gateway with the configured policy (Balanced is the default)
smesh start
# Point any client at the mesh alias path
export OPENAI_BASE_URL="http://localhost:51737/p/openai/v1"
export OPENAI_API_KEY="<your-provider-key>"
Use these commands only after a verified version appears on the download page. An unpublished channel is not an installable release. Download page →
Install it yourself in a minute, or ask us about rolling it out to a team with shared policy.