Bring every agent under management.Dome is GA.

Start free
Dome Systems

Use case

Consolidate tool access

Using Dome's MCP Gateway

Every agent reaches for tools: Zendesk, Jira, a warehouse, three internal APIs, someone's custom MCP server. Consolidating that reach is six resources, and the agent is granted exactly one of them. Credentials stay at the boundary, rules run on every call, and one audit vocabulary covers all of it.

Composed from

Six resources, one boundary

Gateways, Tools, Rules, Guards, Quotas and Audits. The gateway is where the six meet, whether your agents are the ones calling out or your customers' agents are the ones calling in.

Gateways

The one surface an agent is granted. Logical, not physical. Make as many as the shape of your estate needs.

Tools

The connections behind it. An MCP server, a REST API described by OpenAPI, or a first-party internal service.

Rules

Evaluated per call, on the arguments as well as the target. Simulated against real traffic before they ship.

Guards

The response path. Redact fields, refuse records, filter what comes back before it reaches the model.

Quotas

Ceilings per agent, per tenant, per window.

Audits

Every call, with who, what, under which rule, and with what result. One vocabulary across every backend.

Nothing here belongs only to the MCP Gateway. One policy language, one audit vocabulary, one access layer, across every composition. See the whole resource set.

Scenario

Four tools, four credentials

A support-triage agent reaches Zendesk, Jira, Snowflake and a custom MCP server. Four credentials sit in its environment, four auth models sit in its code, and audit comes back in four shapes.

What makes it hard

Credentials travel with the agent

A key in an environment variable is a key on every machine the agent runs on.

The same control, re-implemented per tool

Each backend arrives with its own idea of who may call it and what they may ask for.

Authorization stops at the call

Permitting a query does not constrain the rows it returns.

Cost has no boundary

One conversation fans out into dozens of tool calls. A per-request budget tells you nothing.

The work, in resources

  1. Tools. Add each backend once. An MCP endpoint, an OpenAPI description, or an internal service.

  2. Gateways. Compose them into one Gateway and grant it to the agent. The credential stays here and never reaches agent code.

  3. Rules. Permit the operations this agent needs. Simulate against yesterday's traffic before you deploy them.

  4. Guards. Filter the response path so a permitted call cannot return fields this caller should not see.

  5. Quotas. Cap the agent over a window. A retry loop stops before it drains the workspace.

  6. Audits. One event shape across all four backends. Queryable, and streamable to the SIEM.

Scenario

Your API becomes a tool surface

You publish an MCP surface over your billing API. A customer's agent walks it in a sequence nobody designed for: twenty calls where a person would have made two, half of them retried.

What makes it hard

Callers stopped being deterministic

Rate limits and API keys assume software that behaves the same way twice. An agent decides what to call at runtime.

The contract described none of this

Long-lived sessions, agents spawning agents, one agent holding forty tools at once.

Per-tenant blast radius

One customer's runaway agent is your incident, and containment has to be set in advance.

The arguments are the risk

The call is permitted. What it is asking for is the open question.

The work, in resources

  1. Tools. Expose the operations you want callable. The rest of the API stays invisible, including in discovery.

  2. Gateways. A Gateway per tenant, carrying only what that tenant is entitled to reach.

  3. Rules. Write once at tenant scope, with overrides where a contract genuinely differs.

  4. Guards. Constrain arguments on the request path. A malformed or hostile call is stopped before it runs.

  5. Quotas. Ceilings per tenant. One customer's loop stays one customer's problem.

  6. Audits. Per-tenant usage and denials, ready for billing and for the support conversation.

What changes

Where the operational work ends up

The work does not disappear. Per-tool, per-agent, per-team controls collapse into one place the platform team operates and security can reason about.

Before

After

Distributed control

Credentials live in each agent. Each tool has its own auth model. Audit is per-tool and per-format. The same decision is duplicated across every agent that calls the same backend.

One gateway

Credentials live at the boundary. Policy attaches to the tool. Audit is one event vocabulary across every action. Adding a tool is a configuration change.

Output trust

Whatever the tool returns flows back to the agent. PII, restricted records, credentials embedded in payloads, all of it landing in the model context.

Output governance

Responses pass back through policy. Fields redacted, records refused, sensitive substrings filtered. Outputs governed as well as inputs.