# AgentBack: AI-native API/MCP framework for agents

> Source: <https://agentback.dev>
> Published: 2026-06-15 19:47:13+00:00

### Proven DI core

A hierarchical `Context`

of `Binding`

s with
`@inject`

, providers, interceptors, extension points,
and tag-based discovery — ported from LoopBack 4.

[Dependency injection →](docs/concepts/dependency-injection.html)

Agent-native, end to end

AgentBack gives your REST endpoints, MCP tools, OpenAPI docs, typed
clients, and validation **one Zod contract** — one
schema, every boundary. Agents consume that surface with
machine-actionable errors, `/llms.txt`

, and MCP tools;
and because there's a single source of truth, a coding agent can
author the whole surface from it without drift. Built on
LoopBack 4's proven dependency-injection core, rebuilt for ESM
and Node 22.

alpha · ESM-only · Node 22.13+ · TypeScript 6

One artifact, many views

Declare a schema once, on the decorator. The framework derives every other representation from it — so a coding agent adds a surface in one edit and nothing drifts.

`z.infer`

type`/openapi.json`

, rendered in Swagger
UI
`/llms.txt`

Positioning

Most stacks keep the runtime contract, the service contract, and the agent contract in three hand-synchronized places.

| Stack | Runtime contract | Service contract | Agent/tool contract |
|---|---|---|---|
| Express + raw Zod | Hand-wired Zod | Hand-written OpenAPI | Hand-written manifest |
| tRPC | Zod | TypeScript-only | Custom adapter |
| NestJS | class-validator | Swagger decorators | Custom adapter |
| FastAPI | Pydantic | OpenAPI from same models | Custom adapter |
| AgentBack | Zod | OpenAPI from same Zod | MCP from same Zod |

What's in the box

A hierarchical `Context`

of `Binding`

s with
`@inject`

, providers, interceptors, extension points,
and tag-based discovery — ported from LoopBack 4.

REST and MCP servers are components over the same container. Run either, or both from one process — controllers and tool classes are just bindings.

Stable codes, per-field issues, the violated schema, retryability, and remediation hints — the same machine-actionable envelope on REST and MCP.

`confirm:`

payload-bound confirmation tokens and
`idempotency:`

key replay, declared on the route or
tool that must not run twice.

`@price('$0.001')`

meters a route or tool; the price
gate refuses unpaid calls with an x402/MPP challenge. Stripe
metered billing from the same usage log.

TypeScript consumers import the same schemas and get typed calls plus runtime validation — the client never imports a running server.

Every tool definition costs context-window tokens on every
connection. `toolCostReport`

token-prices the MCP
surface before agents pay for it.

Auth (JWT, OAuth 2.1), authorization voters, health probes, Prometheus metrics, OpenTelemetry, rate limiting, validated config — all as DI components.

Documentation

Three ideas carry the whole framework: everything is a binding in a context, schemas live once on the decorator, and servers are components.

For coding agents

AgentBack ships an
[agent skill](skills/agentback/SKILL.md) that teaches
Claude Code, Codex, Cursor, and 20+ other agents the conventions
that aren't guessable from type signatures — schema-on-decorator,
the slot-0 input bundle, the DI container. One Zod schema is the only
source of truth, so an agent adds a route, its MCP tool, and its docs
in a single edit without drift. The docs are agent-readable too —
[/llms.txt](llms.txt),
[/llms-full.txt](llms-full.txt), and a markdown mirror of
every page — so the agent reads the framework the way it reads your
API.

Built with AgentBack

Not a snippet — a real service. `agentback-demo`

is a
Weather server where one Zod schema set is served over stdio,
authenticated HTTP, and a dev console. It's the shape a coding agent
and the skill produce: one contract fanning out to REST, MCP, a
typed client, and docs, with nothing kept in sync by hand.

One Zod schema set, served three ways — stdio, authenticated HTTP, and a dev console. Backed by the free Open-Meteo API.

`AgentError`

, `/llms.txt`

, tool-cost reportQuick start

AgentBack is in alpha — the end-to-end examples work and the API surface is still moving. Clone the repo and run a working REST, MCP, or hybrid app, each with its explorer UI.