Altrace runs inside the same pod as your agent. That lets you govern every local action that your agents run.
The difference
Current tools sit outside your workload and watch traffic routed to them. Any local actions never reach existing lines of sight. Altrace injects enforcement locally.
Nothing is rerouted manually or through an external hop.
When an agent spawns another agent, the new process inherits the same pod and enforcement boundaries automatically.
LangGraph, CrewAI, AutoGen, or custom loops—in-pod enforcement observes the underlying process rather than relying on framework conventions.
01 · Discover
Surface agents the moment they transmit their first request, with per-agent model and host visibility, alongside an MCP scanner that audits connected tool servers.
Unregistered agents and runtime sub-agents are flagged immediately upon appearance.
Track models, hosts, and tools attributed directly to a team and an agent on every single request.
Audit the MCP tool servers your agents connect to for hardcoded credentials, unsafe permissions, and supply-chain risk.
02 · Enforce
Every request passes through an ordered decision chain before reaching a target, ensuring strict boundaries.
Halt any agent instantly by team or globally. New requests are blocked synchronously and active streams are cancelled.
Requests are cost-checked prior to provider egress. Over-budget requests are dropped before the model is ever called.
Enforce explicit rules and authorization matrices across connected Model Context Protocol tool actions.
03 · Credential Indirection
Agents authenticate to Altrace using scoped virtual keys. Altrace holds the real provider credential, injecting it securely only when an authorized request leaves for the model.
Agents present proxy-issued tokens rather than raw provider keys, preventing leakage or token theft.
Restrict keys to approved models and endpoints so agents cannot exceed their intended operational boundaries.
Revoke a virtual key instantly and refuse subsequent requests while simultaneously triggering the agent kill switch.
04 · Security & Compliance
Enforce foundational network and access boundaries out of the box.
Outbound traffic is encrypted in transit using TLS. Client certificates are verified against your CA with fail-closed production rules.
Restrict control plane operations to authorized networks and trusted proxy headers, blocking unverified inbound requests.
Comprehensive logs capture all decisions, allow/block actions, and identity attributes for seamless SIEM ingestion.
Ecosystem Compatibility
As a transparent, in-pod proxy, Altrace governs traffic on both requests and responses without modifying agent code. Compatible with Anthropic, OpenAI, Azure OpenAI, Google Gemini, AWS Bedrock, self-hosted models (vLLM, Ollama), MCP, LangGraph, CrewAI, AutoGen, and Kubernetes.