Control what agents can access.

Instead of allowing AI agents to call large language model (LLM) APIs directly using hardcoded keys, Guild acts as a centralized control plane and proxy. Every LLM call from an agent is routed through the Guild LLM Gateway. This mediation allows the platform to intercept requests to enforce rate limits, apply content policies, track exact token consumption, and rotate credentials centrally without requiring updates or modifications to the agent’s code.

Guild assigns a distinct, individual identity to every AI agent. Instead of storing production API keys (like an OpenAI or Anthropic secret key) within the agent configuration, permissions are handled within Guild’s control plane using configuration blocks. You define explicitly which workspaces, repositories, resources, and specific API operations (e.g., list_pull_requests, get_commit) an agent identity is allowed to perform.

Because credentials and API keys are managed inside Guild’s centralized control plane rather than being packaged inside individual agent environments, key rotation occurs at the gateway level. When a key is rotated, the change is instant and transparent to the active agents, ensuring continuous operational uptime during security cycles.

When an agent attempts an action flagged as high-risk (such as deploying code, altering production environments, or handling sensitive financial transactions), the control plane pauses execution. The agent generates a proposal for the action, which is routed to a human reviewer. The action cannot execute until explicit human approval is granted through the control plane, ensuring agents cannot autonomously bypass safety boundaries.

Guild records an absolute historical log of every agent execution. This includes every LLM prompt and response, internal tool invocations, logic decisions, and data payloads in real time. These logs provide a complete, immutable audit trail necessary for compliance verification, such as answering security questions during SOC 2 audits.

Guild’s formal SOC 2 certification is currently in progress. To assist organizations with immediate compliance and risk assessment requirements, the platform provides built-in compliance-ready infrastructure, including granular scoped credentials, immutable session logs, full system audit trails, and data mediation.

Guild tracks consumption at the individual resource level. Token usage and financial costs are calculated and mapped directly to specific agents (e.g., distinguishing spend between a PR Reviewer agent vs. an Issue Triage agent), individual users, and specific models (such as GPT-4o, Gemini, or Claude).

Yes. Through the LLM Gateway, administrators can set specific spend thresholds and rate limits on a per-agent and per-model basis. If an agent goes into an execution loop or exceeds its allocated budget, the control plane enforces the rate limit or spend ceiling to block further API consumption, preventing unexpected invoices.