Crate corrosive_agents 

corrosive_agents

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

Highlights

• Builder pattern — assemble an Agent fluently with AgentBuilder.
• JSON manifests — an agent’s name, version, capabilities, skills and MCP servers load from a JSON file (AgentManifest).
• Verifiable identity — Ed25519 signatures over the manifest (identity::AgentIdentity); anyone can verify an agent with its public key.
• NVIDIA NIM client — chat + streaming + embeddings against the OpenAI-compatible integrate.api.nvidia.com endpoint (llm::NvidiaClient).
• Skills & MCP — native async skills::Skills plus a Model Context Protocol stdio client.
• Multi-transport — Tokio REST API + WebSocket (feature server), and gRPC (feature grpc).
• Vector stores — one vector::VectorStore trait; Pinecone (feature pinecone), Qdrant (feature qdrant), an in-memory reference implementation, or your own custom store.

Quickstart

use corrosive_agents::prelude::*;

let agent = Agent::builder()
    .name("research-agent")
    .version("0.1.0")
    .description("Answers research questions")
    .capability(Capability::new("chat", "Conversational Q&A"))
    .llm(NvidiaClient::from_env()?.with_model(models::LLAMA_NEMOTRON_SUPER_49B))
    .generate_identity()
    .build()?;

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

Feature flags

Feature	Default	Enables
server	yes	REST + WebSocket serving (axum)
grpc	no	gRPC serving (tonic, vendored protos)
pinecone	no	Pinecone vector store backend
qdrant	no	Qdrant vector store backend
full	no	All of the above

Re-exports

pub use agent::Agent;

pub use agent::AgentBuilder;

pub use agent::AgentInfo;

pub use agent::AgentManifest;

pub use agent::Capability;

pub use error::Error;

pub use error::Result;

Modules

a2a

Agent-to-agent (A2A) delegation.

agent

The agent core: manifest, capabilities, builder, and runtime.

authgrpc or server

Authentication for agent transports (features server / grpc).

error

Unified error and result types for the crate.

grpcgrpc

gRPC serving (tonic), enabled with feature grpc.

identity

Agent identity: Ed25519 keypairs for signing and verifying agents.

llm

LLM integration: provider traits, chat types, and the NVIDIA NIM client.

mcp

Model Context Protocol (MCP) support.

prelude

Convenient glob-import for the most commonly used types.

serverserver

REST + WebSocket serving (Tokio/axum), enabled with feature server (on by default).

session

Session stores: pluggable persistence for conversation history.

skills

Skills: named, typed, async abilities an agent can execute on demand.

tlstls

TLS configuration helpers (feature tls).

trust

Key rotation, revocation, and trust decisions for agent identities.

tutorial

The step-by-step tutorial: from an empty project to a production-shaped agent. Rendered from docs/TUTORIAL.md; its core code blocks are compile-checked as doctests.

vector

Vector stores for retrieval-augmented agents.

x509x509

X.509 certificate-based identity (feature x509).