Releases·Tool Release
Sentelabs releases extensible-mcp, an open-source proxy enforcing policy between LLMs and MCP servers
Sentelabs cofounder Banu announced extensible-mcp, an open-source proxy that sits between LLMs and MCP servers. It enforces deterministic policy outside the model's reach and limits which tools an agent loads, treating every loaded tool as added context and attack surface. The tool targets prompt-injection risk by restricting agent capabilities to only what policy allows.
2 engagement1 sourceTue, Sep 1, 2026, 11:32 AM
AI Overview
- Cofounder Banu of sentelabs.ai announced extensible-mcp on GitHub as an open-source proxy between LLMs and MCP servers
- Policy enforcement lives outside anything the model can access, since prompts and inputs the agent reads cannot be fully trusted
- The proxy restricts agent capabilities to only what a deterministic policy allows
- Models need not load every available tool; each loaded tool adds context and additional attack surface
Entities
Related
Products·building
Bouncer: MCP proxy that blocks agents from sending data to untrusted destinations
Bouncer is a local MCP proxy that gates outbound tool calls from AI agents, denying calls to destinations derived from untrusted content (like web pages or emails) and allowing only explicitly trusted ones. It uses deterministic Python policy enforcement rather than an LLM in the enforcement path, preventing prompt-injection attacks that trick agents into leaking API keys or other sensitive data. It solves the problem of agents acting on malicious instructions embedded in untrusted content, protecting users and developers building agentic systems.
Bouncer is a local MCP proxy that gates outbound tool calls from AI agents, denying calls to destinations derived from untrusted content (like web pages or emails) and allowing only explicitly trusted ones. It uses deterministic Python policy enforcement rather than an LLM in the enforcement path, preventing prompt-injection attacks that trick agents into leaking API keys or other sensitive data. It solves the problem of agents acting on malicious instructions embedded in untrusted content, protecting users and developers building agentic systems.
6 engagement1 sourceredditSat, Aug 15, 2026, 05:50 AM
Products·launched
toolbroker: open-source middleware for LLM agent tool calls with typed contracts, credential injection, and refusal audit
toolbroker is an open-source Python layer that sits between an LLM agent and its tools. It enforces typed tool contracts, injects credentials so tools never see API keys, runs policy checks, and logs an audit trail that includes refused calls and the reasons. It targets developers building LLM agents who need policy enforcement and auditability beyond simple call logging.
toolbroker is an open-source Python layer that sits between an LLM agent and its tools. It enforces typed tool contracts, injects credentials so tools never see API keys, runs policy checks, and logs an audit trail that includes refused calls and the reasons. It targets developers building LLM agents who need policy enforcement and auditability beyond simple call logging.
2 engagement1 sourceredditMon, Aug 31, 2026, 02:47 PM
Products·launched
Doberman: open-source authorization layer that gates AI coding agent tool calls
Doberman is an open-source authorization layer that sits between a coding agent and its tools. Every tool call passes through a local policy engine that returns allow, authenticate, or block before execution. It targets developers who want to contain agents that go off script and protect their infrastructure.
Doberman is an open-source authorization layer that sits between a coding agent and its tools. Every tool call passes through a local policy engine that returns allow, authenticate, or block before execution. It targets developers who want to contain agents that go off script and protect their infrastructure.
3 engagement1 sourceredditTue, Sep 1, 2026, 08:23 AM
Releases·Tool Release
PolicyAware open-source Python library launches for AI governance
PolicyAware, an open-source Python library and control plane, was introduced to provide governance, security, and compliance for AI applications including LLMs, RAG systems, and AI agents. It evaluates requests before they reach models or external tools, enforcing deny-by-default governance and detecting sensitive data.
PolicyAware, an open-source Python library and control plane, was introduced to provide governance, security, and compliance for AI applications including LLMs, RAG systems, and AI agents. It evaluates requests before they reach models or external tools, enforcing deny-by-default governance and detecting sensitive data.
2 engagement1 sourceredditFri, Jul 31, 2026, 06:00 AM
Products·building
Envoy proxy extension that turns wire protocols into structured statements for LLM analysis
An Envoy proxy extension that parses request traffic at runtime — via a deterministic SQL parser, gRPC payload inspection, and LLM analysis — into structured statements an LLM can reason over. Security teams use it to proxy headless agent traffic and block harmful actions (like unwanted writes) at the network layer before requests reach upstream services.
An Envoy proxy extension that parses request traffic at runtime — via a deterministic SQL parser, gRPC payload inspection, and LLM analysis — into structured statements an LLM can reason over. Security teams use it to proxy headless agent traffic and block harmful actions (like unwanted writes) at the network layer before requests reach upstream services.
4 engagement1 sourcehackernewsThu, Sep 10, 2026, 04:25 PM