Product · MCP Gateway
Hodor's MCP Gateway gives each agent its own non-human identity, constrains tool calls down to the payload level, and records every call into a replayable audit trail — so your team builds with MCP instead of hand-rolling the controls behind it.
What you get
Per-agent identity
Each harness gets its own cryptographically-rooted identity — not a shared OAuth grant, not a service-account key pasted into a config file. Tokens are minted just-in-time, expire in minutes, and the real credentials never touch the model.
Tool- and payload-level policy
Enable or disable tools per agent — the rest are invisible to the model. Then go deeper: constrain arguments, mask PII before the call leaves your network, shape responses before they reach the model. Most misuses simply can’t be expressed.
Replayable audit
An immutable log of who called what, when, against which policy was in force. Replay any decision with the exact policy in effect. Stream to Grafana, query via MCP, export as signed evidence — the audit stands on the trail, not on trust.
Drop-in install
Point any MCP-compatible harness — Cursor, Claude Code, Mastra, CrewAI, AutoGen, a custom build — at the Hodor endpoint with a bearer token. No SDK in the agent. No fork of the MCP server. The gateway slots in.
Built-in guardrails
Tools not in an agent’s policy bundle are never advertised, never resolved, and never invoked. The model can’t discover what isn’t there — and the request never executes if a payload violates a constraint.
Unified observability
Track AI adoption across teams, watch latency by tool and by agent, alert on unexpected verdicts. Pipe events to the observability stack you already run — or query them with the agents you’re already running.
Drop-in install
Point your harness at api.hodor.ai/mcp. Paste a static agent token for the three-line path, or use Agent OAuth so the agent authenticates as its own identity — same endpoint either way, no SDK in the agent.
An external app is requesting access to your Hodor agent tools via MCP.
Select an agent identity
The connecting app can invoke your agent’s tools on your behalf.
Cursor, Claude Code, Mastra, CrewAI, AutoGen, any MCP client — same endpoint.
Workforce access
Plug Hodor into the identity provider you already run. SAML 2.0 and OIDC out of the box, SCIM that mirrors your IdP groups into Hodor groups, and delegated tokens so the gateway can authenticate to the downstream MCP or API through your IdP — as the agent, never as a shared secret.
Manage who can access this workspace and how they're grouped.
| Member | Role | Status | |
|---|---|---|---|
AR Alex RiveraYou alex@northwind.io | Admin | Active | |
PN Priya Nair priya@northwind.io | Admin | Active | |
DC Daniel Cohen daniel@northwind.io | Member | Active | |
LM Léa Martin lea@northwind.io | Member | Active | |
Marco Bianchi marco@northwind.io | Member | Active |
Okta
Microsoft Entra
Google Workspace
Auth0SAML 2.0, OIDC, SCIM · or talk to us for custom IdPs.
Where Hodor goes further
01
Most gateways apply workspace and user permissions — built for humans clicking through a UI. Hodor mints a distinct identity for each agent run and scopes that identity, not a person who happens to be on call.
02
Enable-or-disable on tools is table stakes. Hodor constrains what each tool can be called with — argument shapes, value ranges, fields the model never sees in the response. The expected behavior becomes the only allowed behavior.
03
Not an LLM router with MCP bolted on. The product was designed from day one for a world where agents call tools that change real state.
Bring an agent and one tool that scares you. We'll walk through the policy that catches it.
