Veea Open-Sources Lobster Trap to Monitor AI Agent Conversations
"What has been missing is a practical way to observe and enforce policy at the point where the AI agents interact with AI models. Lobster Trap addresses that gap."
Veea Inc. announced the open-source release of Lobster Trap at the Mobile World Congress 2026 in Barcelona on March 2. The security tool monitors interactions between AI agents and the language models that power them.
Available under the MIT license at github.com/veeainc/lobstertrap, Lobster Trap functions as part of TerraFabric, Veea's platform for managing autonomous systems at the network edge.
Veea also formed a partnership with NativelyAI to integrate Lobster Trap into Native.Builder, a software production platform used by lablab.ai's community of over 250,000 developers.
The integration allows teams to ship AI agents with security controls already configured, rather than adding them after deployment. The partnership aims to build reusable security templates and policy configurations for different industries.
Modern AI agents perform tasks that extend far beyond answering questions. They access files, generate code, send communications, and execute commands within production environments.
When prompts are manipulated or models return unexpected outputs, the consequences can include credential leaks, unauthorized data access, or unintended system actions.
Traditional security infrastructure designed for web traffic and API calls does not provide visibility into the back-and-forth exchanges between agents and models.
Allen Salmasi, Founder and CEO of Veea, framed the problem as a mismatch between capability and oversight.
"The industry has spent the last two years racing to give AI agents more power," Salmasi said. "What has been missing is a practical way to observe and enforce policy at the point where the AI agents interact with AI models. Lobster Trap addresses that gap."
Lobster Trap operates as an inline filter positioned between agents and models. Each request an agent makes and each answer it receives passes through policy checks before execution continues.
If a violation is detected, Lobster Trap can block the interaction, flag it for review, or log it for analysis. The scanning occurs in under a millisecond and introduces no meaningful delay.
The system supports any backend that implements the OpenAI-compatible API format, which means organizations can typically deploy it without rewriting application logic.
Security policies are defined in a configuration file. Out of the box, Lobster Trap detects prompt injection attempts, credential exposure, personal information leakage, suspicious file access, and data exfiltration patterns.
Integration and Partnership
TerraFabric, Veea's broader control system for edge computing, treats distributed hardware as unified systems instead of standalone units. It handles orchestration, enforces policies, and manages software updates across device fleets.
Lobster Trap applies similar governance principles to the AI layer. TerraFabric determines which workloads operate and their deployment locations. Lobster Trap determines what questions those workloads can pose to language models and what responses they can accept.
When deployed through TerraFabric, Lobster Trap executes on edge hardware rather than in cloud environments. This architecture allows prompt content, model outputs, and security logs to stay within organizational boundaries, addressing compliance requirements around data residency.
"Autonomy without control introduces risk," Salmasi said. "TerraFabric governs infrastructure. Lobster Trap governs AI conversation. Together, they give operators policy enforcement at every layer."
The NativelyAI collaboration centers on embedding Lobster Trap into workflows developers already use. NativelyAI operates lablab.ai, a platform where more than 250,000 developers build and test AI applications.
By incorporating Lobster Trap into Native.Builder, the company's production toolchain, Veea aims to make security enforcement a default feature rather than an optional add-on.
The partnership includes work on industry-specific policy templates and reference architectures that teams can adapt to their environments.
This approach reverses the typical sequence where security controls are added after applications reach production. With Lobster Trap integrated into the build process, agents ship with monitoring and policy checks already in place.
Open Source and Availability
Veea released Lobster Trap under the MIT license, enabling developers to use, modify, and extend it freely. The project is written in Go and compiles to a single file with no external dependencies, allowing it to run on Linux, macOS, or Windows.
The lightweight design makes it suitable for deployment in edge environments where resource constraints are a consideration.
Lobster Trap is available now at github.com/veeainc/lobstertrap. TerraFabric is currently in early deployments.
Organizations interested in evaluating TerraFabric with integrated Lobster Trap capabilities can request early access at veea.com.