Anthropic has expanded the capabilities of its Compliance API by introducing new local session transcript endpoints for Claude Code, giving enterprise security teams greater visibility into how AI coding agents operate on developer endpoints. The update addresses one of the key challenges introduced by local AI harnesses, which execute shell commands, access local files, invoke Model Context Protocol servers, and use credentials already available on developers’ machines. According to Token Security, local AI agents now account for 68.6 percent of AI agents discovered in customer environments, highlighting the growing shift from browser based AI assistants to endpoint based autonomous tools. While the new API improves monitoring capabilities, researchers note that activity logs alone are not sufficient to determine whether an AI agent is using its access appropriately. More information about the Compliance API is available at https://www.anthropic.com/, while Token Security has also shared research on AI agents at https://www.token.security/the-agentic-pulse.
Unlike traditional chatbots, Claude Code functions as a harness that orchestrates interactions between developers and large language models while executing commands directly on local machines. The language model processes requests in the cloud, but the harness performs operations such as running Bash commands, accessing files, authenticating with third party services, and connecting to MCP servers. This architecture creates governance challenges because many actions occur outside a centralized cloud environment. Anthropic’s managed settings provide organizations with a policy baseline by allowing administrators to define allow and deny lists for MCP servers, restrict Bash commands through regular expressions, and disable selected skills. These settings can be deployed through enterprise management tools or mobile device management platforms, but researchers note that static policies cannot fully account for execution context, user intent, or changing operational requirements.
The newly introduced Compliance API endpoints provide access to local session metadata and complete session transcripts for Claude Code interactions with Anthropic models. These transcripts record user prompts, tool usage, Bash commands, MCP interactions, file reads and writes, and other operations that are transmitted to the language model. Organizations can use this information to build inventories of AI agents, identify the MCP servers, skills, and plugins they use, and monitor administrative actions. However, researchers explained that some endpoint activities remain outside the scope of the Compliance API. OpenTelemetry continues to play an important role because it captures local events such as permission changes, hooks that intercept tool execution, and actions that never reach the language model. In addition, organizations using Claude Code with models hosted on services other than Anthropic, including Amazon Bedrock, Microsoft Foundry, or Google Cloud, will not receive Compliance API coverage because only interactions with Anthropic models are recorded. Researchers also cautioned that session transcripts may contain sensitive information including credentials, customer information, and personally identifiable information, making secure storage and controlled access essential.
Researchers emphasized that effective governance for local AI agents requires combining multiple layers of visibility rather than relying on a single data source. Endpoint telemetry can reveal configuration files, installed skills, plugins, and local processes that cloud services cannot observe, while endpoint detection and response platforms provide evidence of activity without connecting actions to identity, ownership, or intended purpose. Token Security noted that organizations typically have more than ten configuration files associated with each local AI agent, and many community developed MCP servers operate outside vendor ecosystems. The report concludes that managed settings, Compliance API session transcripts, OpenTelemetry, and endpoint security data collectively improve visibility, but organizations still need additional context to determine whether AI agents are operating with appropriate permissions. Correlating endpoint activity with internal repositories, approved plugins, organizational identities, and least privilege access policies enables security teams to distinguish legitimate development activity from potentially unauthorized behavior and strengthen governance across enterprise AI deployments.
Follow the SPIN IDG WhatsApp Channel for updates across the Smart Pakistan Insights Network covering all of Pakistan’s technology ecosystem.
