Source: https://eisoptric.com/mcp
中文: https://eisoptric.com/zh/mcp.md
Built: 2026-11-20

Eisoptric MCP · A tool from Eisoptric’s research

# The MCP gateway that decides what your agents may do.

Eisoptric MCP is the MCP server your agents talk to. Behind it sit your real MCP servers — CRM, payments, databases, cloud consoles. Eisoptric exposes only the tools you approve, evaluates every call against policy, pauses the sensitive ones for a human, and records everything.

[Get started](https://console.eisoptric.com/) [Read the whitepaper](https://eisoptric.com/trust)

Standards-compliant MCP over HTTP. Works with Claude, ChatGPT, Cursor, Eisoptric Workbench, or any MCP client.

```
policy    Payments  v3  production

rule      payments.transfer
when      amount > 1000 AND currency == "USD"
decision  approval_required

rule      payments.transfer
when      true
decision  allow

rule      db.drop_table
when      true
decision  deny
```

A policy as the engine evaluates it. Conditions are matched against the arguments of each call.

## Up and running in four steps.

No SDK, no agent code changes. Eisoptric works at the protocol level.

01

### Connect your MCP servers

Register each server by URL. Eisoptric discovers its tools, handles OAuth where needed, and seals the credentials in its vault.

02

### Build tool sets

Pick the tools each agent should have — across servers. Mark any tool as Call (automatic) or Approval (needs a person).

03

### Write policies

Add rules with conditions on arguments. Decide per rule: allow, approval required, or deny. Policies are versioned.

04

### Point agents at Eisoptric

Give every agent the same Eisoptric MCP endpoint. They sign in with OAuth and see only their tool set.

## What teams run through Eisoptric.

Four examples of agents that are useful precisely because they are governed.

Finance operations

### An agent that pays vendors — with a human on large amounts

The finance agent reconciles invoices and triggers payments through the payments MCP server. Small transfers run automatically. Anything above the threshold waits for a reviewer, who sees the amount, the recipient, and the invoice reference before approving.

-   payments.transfer · amount ≤ 1000 → allow
-   payments.transfer · amount > 1000 → approval required
-   payments.refund · amount > 10000 → deny

Customer support

### A support agent that can read and reply, but never delete

The support agent gets a tool set with CRM lookup, ticket updates, and email reply. Destructive CRM tools are not in the set, so for this agent they do not exist. Emails to external domains require approval; internal notes go straight through.

-   crm.get\_customer · allow
-   tickets.update · allow
-   email.send · recipient not in company domain → approval required
-   crm.delete\_customer · not in tool set

Platform and DevOps

### A release agent that deploys to staging freely, to production with sign-off

The release agent runs rollouts through your deployment MCP server. Staging is automatic. Production rollouts pause for an on-call engineer. Schema-destroying operations are denied outright, regardless of environment.

-   deploy.rollout · environment == "staging" → allow
-   deploy.rollout · environment == "production" → approval required
-   db.drop\_table · deny

Data and analytics

### Analyst agents for every team, each confined to its own data

Each team gets analyst agents with read-only warehouse tools. Server mapping decides which teams may attach the warehouse at all. Editors are scoped to their own agents, and every query is attributed to the person the agent acted for.

-   warehouse.query · allow
-   warehouse.export · row\_count > 100000 → approval required
-   Marketing accounts · warehouse server not mapped

## What the gateway handles for you.

Everything between the agent’s request and your upstream server.

### Identity

OAuth 2.1 with PKCE for every agent. Each session is bound to a user or an organization; the profile is checked on every request.

### Tool discovery

Register MCP servers by URL. Eisoptric discovers tools, keeps the catalog current, and manages upstream OAuth and token refresh.

### Tool sets

Curated tools per agent, across servers. Only those tools appear in tools/list. Each tool is Call or Approval.

### Policy engine

Ordered, versioned rules with conditions on arguments. Every call records the policy id, version, and matched rule.

### Approval queue

Sensitive calls wait for one reviewer or a group quorum. The agent resumes with gatewayRunId. Approvals expire.

### Credential vault

Upstream tokens and secrets sealed with ML-KEM-1024 and AES-256-GCM. Agents never see them; every read is audited.

### Audit trail

Governance events stay searchable. Stored audit payloads use zero-access encryption and open only in a member’s browser, while the gateway processes plaintext in flight.

### Signed webhooks

Approval granted, rejected, and expired are signed with HMAC-SHA256. A successful delivery does not execute the tool.

## Security built into the request path.

An agent can never obtain more than the result of an authorized call. These are the mechanisms behind that guarantee.

-   ### Post-quantum credential vault
    
    Access tokens, refresh tokens, and client secrets are sealed individually with ML-KEM-1024 key encapsulation and AES-256-GCM, keys derived via HKDF-SHA384. The key seed lives in configuration, never in the database.
    
-   ### No plaintext over the API
    
    Console and client APIs report whether a secret exists — never its value. Every secret read generates a credential\_accessed audit event.
    
-   ### Managed upstream OAuth
    
    Eisoptric discovers the authorization server, performs dynamic client registration where supported, uses PKCE, and refreshes tokens automatically when an upstream returns 401.
    
-   ### Network boundaries
    
    Upstream URLs must use HTTPS (or HTTP on localhost). Private, link-local, and cloud metadata addresses are refused.
    
-   ### Denied means never sent
    
    A denied call never leaves Eisoptric. A held call returns approval\_required with a reference id. A paused agent cannot start a call.
    
-   ### Role-based access control
    
    Four console resources, three access levels each. Preset roles admin, operator, viewer, or custom sets. Editors can be scoped to specific agents.
    

## Why a dedicated MCP gateway?

API gateways govern services. Eisoptric governs agents — a caller that reasons, chooses tools, and needs to be stopped sometimes.

### Direct connection

-   Agent holds every credential
-   Every tool on the server is exposed
-   No approval step
-   No central audit trail

### API gateway

-   Sees routes and methods, not tools
-   No conditions on tool arguments
-   No human approval workflow
-   Designed for services, not agents

### Eisoptric

-   Credentials sealed in the gateway
-   Tool sets and argument-level policy
-   Built-in approval queue
-   One audit trail for every agent

## One console for every agent in your organization.

Eisoptric Console gives security, platform, and operations teams a shared view of what agents can do and what they have done.

### Overview

Active agents, pending approvals, recent denials, and spend at a glance.

### Agents

Register agents, assign tool sets, pause or resume, and scope who may edit them.

### Eisoptric MCP

Connect MCP servers, review discovered tools, and build tool sets.

### Policy

Author and version rules; see which policy decided each call.

### Intervention

Approve or reject a call. Multi-approver tasks wait until every required group is done.

### Resubmit

Approved tasks wait on their own page until the agent calls the tool again with gatewayRunId.

### Audit

Search governance events. Raw requests and responses open only in an authorized member’s browser.

### Webhooks

Signed notices for approval granted, rejected, and expired. Delivery does not execute the tool.

### Identity

Members, roles, custom permissions, and MCP server mapping per employee.

## Put Eisoptric in front of your first MCP server today.

Create a free account, connect a server, and write one rule. The first approval request shows you the whole path.

[Get started](https://console.eisoptric.com/) [Read the whitepaper](https://eisoptric.com/trust)
