# Agents on your own stack

> AI agents built on your own stack: MCP servers over your REST or GraphQL APIs, read access to your database, jobs on your queue, approvals in your own tools.

Canonical page: https://aispeedforce.com/platforms/custom/

Last updated: 2026-09-23

Not every business runs on a platform. If your work lives in your own application, APIs and databases, we build the agent against those directly, through MCP servers that expose exactly the operations its task needs and nothing more.

## Three shapes we build on custom stacks

Each one is scoped to a single job and reaches your systems through tools we write against your API, not through broad database access.

### Internal operations

Works the queue of manual tasks your ops team handles in an internal admin, calling your own API endpoints to do each step.

- **Touches**: Your REST or GraphQL endpoints, through an MCP server
- **Human gate**: Every write, until you widen it per endpoint.

### Support and account questions

Answers customer and account questions from your data and docs, with read-only access to a replica or reporting database.

- **Touches**: Read replica, docs, ticket system
- **Human gate**: Anything sent to a customer.

### Data and reporting

Pulls figures from your systems on a schedule, checks them against rules you define, and writes a report or raises an alert.

- **Touches**: Warehouse or database reads, your job queue
- **Human gate**: Alerts sent outside your team.

## What it connects to

Five surfaces, defined by you. We write the contract between the agent and your systems, and the agent can do nothing that contract does not name.

- **Your API**: Your existing REST or GraphQL API is the preferred surface, because it already carries your validation and permissions.
- **MCP server**: We build an MCP server over your API with one tool per operation, typed inputs, and clear errors, so any agent runtime can use it.
- **Database**: Read access only, ideally to a replica or reporting database, for questions the API cannot answer.
- **Queue**: Long or scheduled work runs as jobs on your queue, so it retries safely and does not block your application.
- **Internal tool**: The approval queue sits in the internal tool your team already uses, or in a small screen we add to it.

### What each agent reads and writes

Read access is broad enough to answer the question, write access is narrow, and every write has a gate until you decide otherwise.

| Agent | Reads | Writes | Human gate |
|---|---|---|---|
| Internal operations | Your API reads | Your API endpoints named in the contract | Every write, until widened |
| Support and account questions | Read replica, docs | Drafted replies | Anything sent to a customer |
| Data and reporting | Warehouse or database reads | Reports, alerts | Alerts sent outside your team |

## The gate lives in your tools

There is no platform admin to borrow, so the approval step goes where your team already works: your internal tool, or a small queue we add to it.

**Where it sits**: In your own internal tool or admin, as a queue of held actions, with the same login and permissions as the rest of it.

**What a reviewer sees**: The call the agent wants to make, with its inputs, the record it targets, and the run trace behind it.

**What waits by default**: Every write to your systems, until you widen the agent's authority endpoint by endpoint, and anything sent to a customer.

**What a rejection does**: The run stops, the rejection is kept with its trace, and it becomes an eval case the agent is tested against on every change.

## Questions teams with their own stack ask first

### Do you need direct access to our production database?

No. We prefer your API, because it already enforces your rules. Where the agent needs to read data the API does not expose, we ask for a read replica or a reporting database.

### Which languages and frameworks do you work with?

The agent talks to your systems over HTTP or through an MCP server we build, so your application can be written in anything. The MCP server is written to fit your stack and handed over with the code.

### Who owns the MCP server and tools?

You do, on full payment, as set out in the terms. The code, prompts, tool definitions and eval cases live in your repository.

### Where to go next

- [Agent engineering](https://aispeedforce.com/services/agent-engineering/)
- [Automation and workflows](https://aispeedforce.com/services/automation-workflows/)
- [Retrieval and data](https://aispeedforce.com/services/retrieval-and-data/)
- [All platforms](https://aispeedforce.com/platforms/)
- [Contact](https://aispeedforce.com/contact/)

## Name one job in your own systems

Tell us the task your team does by hand in your internal tools. We will tell you which endpoints the agent would need, where the gate sits, and what the MCP server would expose.