corrosive_agents

Build verifiable, interactive AI agents in Rust, powered by NVIDIA Nemotron free LLM models.

A corrosive agent has a name, a version, and a set of active capabilities — all loadable from a JSON manifest — plus native Skills, MCP (Model Context Protocol) servers, and an Ed25519 identity so anyone can verify it with public-key cryptography. Serve it over a Tokio REST API, WebSocket, or gRPC, and give it memory with Pinecone, Qdrant, or any custom vector store.

Features

📖 New to the library? Read the tutorial — it walks from an empty project to a production-shaped agent, and ships on docs.rs as the tutorial module.

Installation

[dependencies]
corrosive_agents = "0.0.1"          # REST + WebSocket by default
tokio = { version = "1", features = ["full"] }

Optional features:

corrosive_agents = { version = "0.0.1", features = ["full"] } # everything

Quickstart

Get a free API key at build.nvidia.com and export it as NVIDIA_API_KEY.

use corrosive_agents::prelude::*;

#[tokio::main]
async fn main() -> Result<()> {
    let agent = Agent::builder()
        .name("research-agent")
        .version("0.1.0")
        .description("A concise research assistant")
        .system_prompt("Answer in at most three sentences.")
        .model(models::NEMOTRON_3_NANO_30B)
        .capability(Capability::new("chat", "Conversational Q&A"))
        .llm(NvidiaClient::from_env()?)
        .generate_identity()          // manifest is signed at build()
        .build()?;

    let reply = agent.chat("session-1", "What is NVIDIA Nemotron?").await?;
    println!("{reply}");
    Ok(())
}

Load an agent from JSON

{
  "name": "manifest-agent",
  "version": "1.0.0",
  "model": "nvidia/llama-3.3-nemotron-super-49b-v1",
  "capabilities": [{ "name": "chat", "description": "Conversational Q&A" }],
  "skills": ["word_count"],
  "mcp_servers": [
    { "name": "fs", "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"] }
  ]
}

use corrosive_agents::prelude::*;
# fn skill_impl() -> FnSkill { FnSkill::new("word_count", "", |i| async move { Ok(i) }) }

# async fn run() -> Result<()> {
let agent = AgentBuilder::from_json_file("agent.json")?
    .skill(skill_impl())              // implementations bind to declared names
    .llm(NvidiaClient::from_env()?)
    .generate_identity()
    .build()?;

agent.connect_mcp_servers().await?;   // spawn + handshake declared MCP servers
let tools = agent.mcp_tools("fs").await?;
# Ok(())
# }

Verify an agent with public-key cryptography

use corrosive_agents::prelude::*;

# fn main() -> Result<()> {
let agent = Agent::builder()
    .name("trusted").version("1.0.0")
    .generate_identity()
    .build()?;

// Ship the manifest anywhere as JSON…
let json = agent.manifest().to_json()?;

// …and anyone can verify it, offline:
let received = AgentManifest::from_json(&json)?;
received.verify()?;                            // embedded public key
received.verify_with(&agent.public_key().unwrap())?; // or a pinned key
# Ok(())
# }

Tampering with any signed field makes verification fail.

Serve it

use std::sync::Arc;
use corrosive_agents::prelude::*;

# async fn run() -> Result<()> {
# let agent = Agent::builder().name("a").version("1.0.0").build()?;
let agent = Arc::new(agent);

// REST + WebSocket (feature `server`, on by default)
agent.clone().serve("0.0.0.0:8080".parse().unwrap()).await?;

// gRPC (feature `grpc`)
// agent.serve_grpc("0.0.0.0:50051".parse().unwrap()).await?;
# Ok(())
# }

REST endpoints: GET /health, GET /agent, GET /agent/manifest, GET /capabilities, GET /skills, POST /skills/{name}, POST /chat, POST /verify, and a WebSocket at /ws with optional streamed chunks.

gRPC service (proto/agent.proto): GetInfo, Chat, ChatStream (server streaming), ExecuteSkill — a generated Rust client ships with the crate at corrosive_agents::grpc::pb::agent_service_client::AgentServiceClient.

RAG with a vector store

use corrosive_agents::prelude::*;
use serde_json::json;

# async fn run() -> Result<()> {
let nvidia = NvidiaClient::from_env()?;
let agent = Agent::builder()
    .name("rag").version("0.1.0")
    .llm(nvidia.clone())
    .embeddings(nvidia)                       // NvidiaClient embeds too
    .vector_store(InMemoryVectorStore::new()) // or QdrantStore / PineconeStore
    .build()?;

agent.remember("Nemotron models are free at build.nvidia.com", json!({"topic": "nvidia"})).await?;
let hits = agent.recall("where are Nemotron models hosted?", 3).await?;
# Ok(())
# }

Bring your own store by implementing the VectorStore trait (three async methods: upsert, search, delete).

Examples

All LLM examples need NVIDIA_API_KEY (or NVIDIA_KEY in a .env file).

Nemotron models

Constants in corrosive_agents::llm::models, all usable with a free key:

• nvidia/nemotron-3-ultra-550b-a55b, nvidia/nemotron-3-super-120b-a12b, nvidia/nemotron-3-nano-30b-a3b (default)
• nvidia/llama-3.1-nemotron-ultra-253b-v1, nvidia/llama-3.3-nemotron-super-49b-v1, nvidia/llama-3.1-nemotron-nano-8b-v1, nvidia/nemotron-mini-4b-instruct
• Embeddings: nvidia/nv-embedqa-e5-v5, nvidia/llama-nemotron-embed-1b-v2

The catalog evolves — list what your key can reach with GET https://integrate.api.nvidia.com/v1/models.

Development

cargo test --features full          # unit + integration + doc tests
cargo clippy --all-targets --features full
cargo fmt --check
cargo deny check                    # advisories, licenses, bans, sources (deny.toml)
cargo llvm-cov --features full      # code coverage report
cargo run --manifest-path tools/protogen/Cargo.toml   # regen gRPC code after editing proto/

License

Licensed under either of

• Apache License, Version 2.0 (LICENSE-APACHE)
• MIT license (LICENSE-MIT)

at your option.