CSL-Core
An open-source command-line tool that adds formal-verification-based security policies to tool calls made by LangChain agents. It operates by inserting a policy guard layer before each tool invocation, enabling security control through formal policy enforcement. The tool supports the Z3 theorem prover for policy verification.
Timeline (last 90 days)
CSL-Core is an open-source command-line tool that adds a formally verified policy layer before each tool call without requiring any changes to existing prompts.
Running the cslcore venom command scans the code repository in read-only mode, identifying @tool-decorated functions, StructuredTools, LangGraph nodes, as well as OpenAI/Anthropic tool schemas, MCP servers, and webhooks.
CSL-Core sets per-call limits and approval thresholds for finance-related tools, applies a whitelist mechanism for tools that execute commands, and restricts write operations to a defined folder scope.
CSL-Core uses the Z3 theorem prover to verify rules before policies take effect, ensuring the rules are mathematically satisfiable and free of contradictions.
CSL-Core adds an additional decorator beneath LangChain's @tool decorator; the tool retains its original name, docstring, and parameter schema, so LangChain still sees the same tool.
CSL-Core's cslcore wire --undo command can restore files exactly as they were.
CSL-Core tools start in log mode by default, without blocking any calls, but recording each call as either allow or would-block.
The cslcore venom --check --fail-on-new-reach command causes the build to fail when someone adds a tool that can move money, execute commands, or publish externally without a corresponding rule in place.
CSL-Core can be installed via pip install csl-core, used by running cslcore setup in the repository, and all data remains local and is never transmitted externally.
For pure schema tools, CSL-Core cannot use a decorator and requires manually adding a single line guard.check(...).
1 more timeline events
All Facts (11)
CSL-Core is an open-source command-line tool that adds a formally verified policy layer before each tool call without requiring any changes to existing prompts.
50%UnverifiedRunning the cslcore venom command scans the code repository in read-only mode, identifying @tool-decorated functions, StructuredTools, LangGraph nodes, as well as OpenAI/Anthropic tool schemas, MCP servers, and webhooks.
50%UnverifiedCSL-Core sets per-call limits and approval thresholds for finance-related tools, applies a whitelist mechanism for tools that execute commands, and restricts write operations to a defined folder scope.
50%UnverifiedCSL-Core uses the Z3 theorem prover to verify rules before policies take effect, ensuring the rules are mathematically satisfiable and free of contradictions.
50%UnverifiedCSL-Core adds an additional decorator beneath LangChain's @tool decorator; the tool retains its original name, docstring, and parameter schema, so LangChain still sees the same tool.
50%UnverifiedCSL-Core's cslcore wire --undo command can restore files exactly as they were.
50%UnverifiedCSL-Core tools start in log mode by default, without blocking any calls, but recording each call as either allow or would-block.
50%UnverifiedThe cslcore venom --check --fail-on-new-reach command causes the build to fail when someone adds a tool that can move money, execute commands, or publish externally without a corresponding rule in place.
50%UnverifiedCSL-Core can be installed via pip install csl-core, used by running cslcore setup in the repository, and all data remains local and is never transmitted externally.
50%UnverifiedFor pure schema tools, CSL-Core cannot use a decorator and requires manually adding a single line guard.check(...).
50%UnverifiedCSL-Core's design philosophy of 'shadow-run first, then enforce' is similar to the detection/blocking mode switching of a WAF.
50%