Nerva — Composable Agent Primitives

Every agent framework wants to own your stack. Nerva doesn’t. It gives you 8 composable primitives — routing, runtime, tools, memory, policy, and more — that drop into the web framework you already use. No lock-in, no opinions about your HTTP layer, no hidden abstractions. Build your agent system, not the plumbing around it.

Library, not framework

Nerva runs inside your web framework. It does not own your HTTP layer, your auth, or your API docs. You bring FastAPI, NestJS, or Express — Nerva handles the agent infrastructure underneath.

Getting Started

Pick your language and build your first agent in 5 minutes.

Python pip install otomus-nerva — build your first agent in 5 minutes.

TypeScript npm install @nerva/sdk — full type safety, streaming-first.

Go Coming soon — join the discussion on GitHub.

The 8 Primitives

Everything you need, nothing you don’t. Each primitive is independent, composable, and replaceable.

ExecContext Request identity, permissions, memory scope, and tracing — threaded through every primitive.

Router Classify intent and select the right handler. Embedding, LLM, rule-based, or hybrid strategies.

Runtime Execute agent code with isolation, timeouts, circuit breakers, and streaming support.

Tools Discover, sandbox, and invoke external tools. MCP servers, functions, composites.

Memory Tiered storage — hot (session), warm (episodes and facts), cold (vector search).

Responder Format agent output for the target channel and tone.

Registry Unified catalog of agents, tools, and plugins. Register, discover, health-check.

Policy Declarative safety, permissions, rate limits, cost budgets, and approval gates.

Use what you need

You don’t need all 8. Use one primitive, use all of them, or replace any with your own implementation. Every primitive has a clean interface you can swap.

How It Works

FastAPI / NestJS / Express   (HTTP, auth, sessions, swagger)
  └── Nerva                  (routing, runtime, tools, memory, policy)
       └── LLM providers, MCP servers, subprocess agents

What’s in the Box

Each primitive ships with multiple built-in strategies. Pick the one that fits, or write your own.

4 Router Strategies

Embedding similarity, LLM classification, rule-based pattern matching, and hybrid (combine any).

3 Runtime Modes

In-process, subprocess isolation, and container sandboxing — with circuit breakers on all three.

3 Tool Managers

MCP server pools, plain function registries, and composite tool chains.

3 Memory Tiers

Hot (session state), warm (episodes and extracted facts), cold (vector search over history).

3 Responder Types

Plain text, structured JSON, and streaming — each with channel-aware formatting.

3 Registry Backends

In-memory, file-based, and remote (HTTP/gRPC) — all behind the same interface.

3 Policy Engines

Static rules, OPA integration, and LLM-as-judge — stack them for defense in depth.

Scaffold in seconds

npx nerva new my-agent --lang python

The CLI generates a working project with your chosen primitives and a test suite. Supports --lang python, --lang typescript, and --template for common patterns.

Next Steps

Architecture How the primitives compose, data flow, and extension points.

Cookbook Recipes for common patterns — multi-agent, RAG, human-in-the-loop, and more.

Migration Guide Moving from LangChain, CrewAI, or a custom setup? Start here.