Argentor

Secure, high-performance multi-agent AI framework in Rust — WASM sandboxed plugins, MCP native, compliance-ready.

Quick Start

git clone https://github.com/fboiero/Argentor.git
cd Argentor
cargo run --example hello_world

Why Argentor?

• Security-first — WASM sandbox + capability-based permissions on every skill; shell injection, null-byte, URL-decode, and NFKC path attacks blocked in the guardrails pipeline
• Rust performance — ~2 ms framework overhead per turn vs 11–55 ms for Python-based competitors (N=10 paired, p < 0.0001)
• WASM sandboxing — plugins run in wasmtime/WASI isolation; ed25519 signature verification and static analysis before load
• MCP native — full client/server/proxy with centralized control plane, credential vault, and token pool

Benchmarks

Argentor vs LangChain, CrewAI, PydanticAI, and Claude SDK across 6 dimensions (reproducible from benchmarks/):

Full report: docs/BENCHMARK_SYNTHESIS.md — includes sensitivity analysis and honest losses.

Architecture

17 crates, strict workspace lints (unwrap_used, expect_used → warn in CI):

Features

Security

• WASM sandboxed plugins (wasmtime + WASI) — capability-based permissions per skill
• Guardrails: shell injection, base64-decode attacks, unicode normalization (NFKC), null-byte, URL-decode path attacks
• SSRF prevention — blocks localhost, link-local, and private ranges
• Path traversal protection — canonicalization + blocklist + null-byte detection
• PII detection (Luhn, SSN, email, phone) with redaction
• Prompt injection blocking — 23+ pattern signatures
• RBAC policy engine, JWT/mTLS, encrypted credential store (AES-256-GCM)
• Audit log rotation with background writer thread

Intelligence

• Token budget per session + adaptive context compaction (4 strategies)
• Extended Thinking Mode (Quick/Standard/Deep/Exhaustive) with confidence scoring
• Self-Critique Loop (Reflexion pattern across 6 quality dimensions)
• Dynamic Tool Discovery (TF-IDF + keyword hybrid, ~98% token reduction)
• Learning Feedback Loop — tool selector improves from execution outcomes
• 6-phase competitive benchmark suite (vs LangChain, CrewAI, PydanticAI, Claude SDK)

Operability and Observability

• REST audit plane: /api/v1/audit/logs, /api/v1/audit/violations, /api/v1/audit/stats with cursor pagination and bounded reads
• Audit dashboard at /dashboard/audit with outcome and date-range filters, JSON detail drawer, and CSV/JSON export
• Persisted audit.jsonl.violations.idx and audit.jsonl.stats.idx keep both endpoints near-constant time across process restarts and 10M-event logs
• SSE session streaming at /api/v1/stream/{session_id} for token, tool call, and done events
• Prometheus audit metric family (argentor_audit_configured, _log_bytes, _events_total, _events_today, _violations_today, _block_rate_percent) plus the existing HTTP/skill/LLM counters
• Webhook event notifications (AgentStarted, AgentCompleted, AgentFailed, ToolCalled, GuardrailBlocked, ApprovalRequired, BudgetExhausted) with HMAC-SHA256 signing
• Audit-scale benchmark suite (cargo run -p argentor-benchmarks -- audit-scale) with recorded 100k/1M/10M latency baselines

Developer Experience

• 5 runnable examples (see examples/)
• Workflow DSL — TOML-based agent workflows, no Rust code required
• Tool Builder — 3-line tool definitions
• Config hot-reload via file watcher (500ms debounce)
• CLI REPL with 12 commands for interactive agent debugging
• Observability dashboard at /dashboard — pure HTML/JS, no build step
• Enterprise readiness report at /api/v1/enterprise/readiness — runtime score, active checks, available controls, and next actions
• Four ready-to-adapt agent profile templates under templates/ (code-reviewer, customer-support, data-analyst, rag-agent)
• Kubernetes Helm chart under deploy/helm/argentor/ with hardened pod security defaults and HPA support

Integrations

• 14 LLM providers: Claude, OpenAI, Gemini, OpenRouter, Groq, Ollama, Mistral, xAI, Azure OpenAI, Cerebras, Together, DeepSeek, Cohere, HuggingFace
• AWS Bedrock backend — real SigV4 signing, feature-gated (bedrock feature flag)
• 5,800+ tool integrations via MCP ecosystem
• Channel adapters: Slack, Discord, Telegram, Webchat
• Google A2A interoperability protocol
• Python SDK (pip-installable) + TypeScript SDK (npm)

Getting Started

Prerequisites

• Rust 1.80+ (rustup update stable)
• An API key from any supported LLM provider

Run Examples

# Hello world agent
cargo run --example hello_world

# Full DevOps pipeline demo (no API key needed)
cargo run -p argentor-cli --example demo_pipeline

# See all examples
ls examples/

Configure

# argentor.toml
[model]
provider = "deepseek"
model_id = "deepseek-chat"
api_key = "${DEEPSEEK_API_KEY}"
max_tokens = 4096
max_turns = 20

[server]
host = "0.0.0.0"
port = 3000

Run the Server

cargo run --bin argentor -- serve
cargo run --bin argentor -- skill list

Run Tests

cargo test -p argentor-core -p argentor-agent -p argentor-security
cargo test --workspace
cargo clippy --workspace   # 0 warnings
cargo fmt --all -- --check

Python SDK

pip install -e python/

from argentor import ArgentorClient

client = ArgentorClient(base_url="http://localhost:3000")
result = client.run_task("my_skill", "input text here")
print(result["response"])

Async support and 24 typed models included. For native Rust bindings: pip install maturin && cd crates/argentor-python && maturin develop.

Observability Dashboard

cargo run --bin argentor -- serve
# Operator cockpits
open http://localhost:3000/dashboard         # deployments, agents, health
open http://localhost:3000/dashboard/audit   # audit log explorer with filters and exports
# Machine-readable surfaces
curl http://localhost:3000/metrics            # Prometheus
curl http://localhost:3000/openapi.json       # OpenAPI 3.0 spec
curl http://localhost:3000/api/v1/audit/stats # JSON summary

Pure HTML/JS — no build step, dark-themed SPA. The audit dashboard adds outcome/date filters, a JSON detail drawer, and one-click CSV/JSON export over the same /api/v1/audit/* endpoints used by external tooling.

Contributing

See CONTRIBUTING.md for guidelines, and CODE_OF_CONDUCT.md.

Bug reports, feature requests, and pull requests are welcome. For security issues, see SECURITY.md.

License

GNU Affero General Public License v3.0 only — see LICENSE.

AGPL-3.0-only applies to the entire Argentor family. No MIT, no Apache dual-licensing.