Platform
Visibility and control
for every AI agent.
Your agents are pushing code, posting messages, and calling APIs on your behalf. Sorin is the proxy between them and your systems — authenticating every agent, enforcing permissions on every action, and logging everything to a single audit trail.
The gap in your security stack.
You have firewalls, CSPM, endpoint detection. None of them know that an AI agent just pushed a config change to your production repo at 3am. The tooling to govern what agents do — not what models say — doesn't exist in most stacks.
Agents act, they don’t just answer
Your coding agent pushes to GitHub. Your support agent posts in Slack. Your ops agent calls internal APIs. These aren’t chatbots — they’re executing real operations on real systems, and most teams have no controls between the agent and the API.
No record of what happened
An agent read a file, created a branch, and opened a PR at 2am. Who authorized it? What model was it running? What was the reasoning? Most teams can’t answer any of these questions because there’s no logging at the agent layer.
Shared keys, broad permissions
Agents get the same API keys as the humans who set them up. One leaked agent key means rotating credentials for the entire team. There’s no way to scope an agent to only the actions it actually needs.
What Sorin does.
Sorin is a proxy that sits between your agents and the APIs they call. Every request passes through — verified, authorized, and logged — before anything touches your systems. It works with MCP-connected coding tools, the Python SDK, or direct HTTP.
See everything
Every agent, every action, every system.
Sorin logs every tool call with the agent identity, action, resource, timestamp, model, and token usage. The activity dashboard shows exactly what each agent did, when, and on which system — with full call-chain context so you can trace a sequence of actions back to the original request.
- Real-time activity feed with full action context
- Per-agent ownership and permission mapping
- Call-chain tracking across multi-step agent workflows
- Execution graph visualization per request
- Session-level grouping of agent actions
Control access
Least-privilege, enforced at the proxy.
Each agent gets its own scoped key and an explicit set of permissions: which connectors it can use, which actions it can take, and whether those actions require human approval. Sorin enforces this at the proxy layer, so policy can’t be bypassed by agent code. Upstream credentials are encrypted and never exposed to agents.
- Per-agent least-privilege permissions
- Human-in-the-loop approval workflows for write actions
- Credential isolation — agents never see upstream API keys
- Per-connector, per-action permission granularity
- Slack notifications on approval requests
Audit everything
Full trail, queryable from the dashboard.
Every proxied request is logged to an audit trail with the agent, action, resource, status, and reasoning context. When something goes wrong, you can trace exactly what happened — which agent, which tool call, which upstream API — instead of piecing together logs from five different systems.
- Structured audit log with agent, action, and resource
- Full causal traceability across chained agent actions
- Model and token usage recorded per LLM call
- Blocked action logging with denial reasons
- Dashboard-queryable by agent, connector, or time range
What this means for your team.
Sorin is built so that using AI agents doesn't mean giving up visibility or control. Here's what that looks like in practice.
Answer questions about your agents
Which agents exist? What can they access? What did they do last night? Sorin gives you a single dashboard where every agent action is logged with full context.
Stop sharing upstream credentials
Register your API keys in Sorin once. Agents authenticate with scoped Sorin keys. If one leaks, revoke it without rotating credentials for the rest of your team.
Let agents ship without losing control
Engineering teams create agents and define what they need access to. Approval gates catch high-risk actions before they execute. Everyone moves faster because the guardrails are built in.
One proxy for every agent type
Claude Code over MCP, custom Python agents via the SDK, direct HTTP from any language. Same permissions, same audit trail, same approval flow — regardless of how the agent connects.
Take control of your AI agents.
See what your agents are doing. Decide what they're allowed to do. Ship with confidence.