Universal AI Agent Orchestration Layer
Cognitive orchestration for AI agents. 14 sisters. 1,395 tests. Zero-compromise safety.
Hydra is the orchestration layer for the Agentra ecosystem. It coordinates 14 specialized AI "sisters" (memory, vision, codebase, identity, and more) through a 5-phase cognitive loop, routes tasks to the right LLM, enforces safety gates, tracks every action with cryptographic receipts, and runs on 256 MB RAM.
- Cognitive Loop — PERCEIVE > THINK > DECIDE > ACT > LEARN with real LLM integration
- 14 Sister Integration — Memory, Vision, Codebase, Identity, Time, Forge, Evolve, Cognition, Reality, Veritas, Aegis, Rail, ACP, CLI
- Multi-Model Routing — Anthropic Claude, OpenAI, Ollama (works offline)
- 6-Layer Security Gate — Perimeter, authentication, authorization, resource limits, data isolation, audit
- Kill Switch — 3-level emergency stop: instant halt, graceful stop, freeze/resume
- Approval Flow — Pause runs for human approval with configurable timeout
- Action Compilation — Zero-token execution for learned patterns
- Skill Fabric — Extensible, sandboxed skill system with MCP and OpenClaw adapters
- P2P Federation — Multi-agent collaboration, skill sharing, delegated tasks
- Animus Prime — AI-native internal language with 6-target compiler (JS, Python, Rust, Go, SQL, Shell)
- Receipt Ledger — Hash-chained audit trail for every action
- 15 Advanced Capabilities — Resurrection, Dream State, Shadow Self, Token Minimizer, Future Echo, Mutation, Forking, and more
- Real-Time Pulse — Tiered response system with prediction and proactive suggestions
- Desktop App — Dioxus native app with Siri-style globe, voice commands
- Observability — Structured logging, metrics, distributed tracing, configurable exports
# From source (recommended)
git clone git@github.com:agentralabs/hydra.git
cd hydra && cargo build --release
# CLI only
cargo install --path crates/hydra-cli# Required: at least one LLM provider
export ANTHROPIC_API_KEY="sk-ant-..."
# OR
export OPENAI_API_KEY="sk-..."# Start the server
cargo run -p hydra-server
# Send a task (in another terminal)
curl -X POST http://localhost:7777/rpc \
-H "Content-Type: application/json" \
-d '{"jsonrpc":"2.0","id":1,"method":"hydra.run","params":{"intent":"Create a REST API for user management"}}'
# Watch events in real-time
curl -N http://localhost:7777/events# Development
cd crates/hydra-native
cargo tauri dev
# Release build
cargo tauri build
# Binary at: target/release/bundle/
| Crate | Purpose |
|---|---|
hydra-core |
Core types, errors, cognitive phase definitions |
hydra-runtime |
Cognitive loop, boot sequence, config, shutdown |
hydra-server |
JSON-RPC 2.0, SSE streaming, HTTP endpoints |
hydra-cli |
Command-line interface |
hydra-native |
Dioxus desktop app with globe UI |
hydra-model |
LLM providers (Anthropic, OpenAI, Ollama), circuit breaker |
hydra-gate |
6-layer execution gate, risk assessment, kill switch |
hydra-sisters |
MCP bridge infrastructure for 14 sisters |
hydra-skills |
Skill registry, sandbox, executor, adapters |
hydra-pulse |
Real-time tiered response, prediction, proactive engine |
hydra-compiler |
Action compilation, pattern detection, routing |
hydra-capabilities |
Resurrection, Dream, Shadow, Minimizer, Future Echo, Mutation, Forking |
hydra-federation |
P2P peers, sync, delegation, skill sharing |
hydra-mcp |
Universal MCP adapter, protocol, transport, schema validation |
hydra-observability |
Structured logging, metrics, traces, exports |
hydra-belief |
Belief tracking and Bayesian updating |
hydra-kernel |
Low-level cognitive kernel, budget, checkpoints |
hydra-ledger |
Hash-chained receipt audit trail |
hydra-db |
SQLite persistence layer |
animus |
AI-native language: AST, script parser, 6-target compiler |
# ~/.hydra/config.toml
[server]
port = 7777
log_level = "info"
[llm]
provider = "anthropic"
[limits]
token_budget = 100000
max_concurrent_runs = 10
approval_timeout_secs = 300
[voice]
enabled = falseEnvironment variables override config file values. See Configuration Guide for all options.
- Quick Start Guide — Get running in 5 minutes
- Configuration — Config file and env vars
- API Reference — JSON-RPC methods and SSE events
- Architecture — System design overview
- Troubleshooting — Common issues and fixes
- CLI Reference — Command-line usage
- Voice Commands — Voice interface guide
- Skill Development — Building custom skills
- Federation — P2P multi-agent setup
- Offline Mode — Running without internet
- Capabilities — 15 capabilities explained
- Animus Prime — AI-native language guide
# Run all tests (1,395 tests)
cargo test --workspace
# Run specific crate
cargo test -p hydra-runtime
# Run with real LLM (requires API key)
cargo test --workspace --features live-llm
# Run with real sister connections
cargo test -p hydra-sisters --features live-sistersHydra includes 15 advanced capabilities:
| Capability | Description |
|---|---|
| Resurrection | Resume from any checkpoint, full state recovery |
| Dream State | Background simulation and scenario exploration |
| Shadow Self | Parallel validation of decisions before execution |
| Token Minimizer | Aggressive context compression |
| Future Echo | Outcome prediction with confidence scoring |
| Mutation | Self-evolving patterns with A/B testing |
| Forking | Parallel universe exploration for decisions |
| Action Compilation | Zero-token execution for learned patterns |
| Skill Fabric | Extensible sandboxed skill system |
| Pulse | Real-time tiered response with prediction |
| Federation | P2P multi-agent collaboration |
| Animus Prime | AI-native language with 6-target compiler |
| Belief Tracking | Bayesian belief updating |
| Receipt Ledger | Cryptographic audit trail |
| Cognitive Loop | 5-phase autonomous reasoning |
| Profile | Token Budget | Concurrent Runs | Memory |
|---|---|---|---|
minimal |
10,000 | 2 | 256 MB |
standard |
100,000 | 10 | 1 GB |
performance |
500,000 | 50 | 4 GB |
unlimited |
No limit | 100 | 8 GB |
See CONTRIBUTING.md for guidelines.
MIT — See LICENSE
Built by Agentra Labs