Arkavo Edge

Instant, secure orchestration for AI agents—launch, mesh, and monitor in real time.

Quick Start

Install on macOS

Download the installer from the releases page, open the .pkg file, and follow the installation wizard.

For advanced users: Install via Homebrew

brew tap arkavo-org/homebrew-arkavo
brew trust --formula arkavo-org/arkavo/arkavo  # required on Homebrew 5.2+
brew install arkavo

Install on Linux

brew tap arkavo-org/homebrew-arkavo
brew trust --formula arkavo-org/arkavo/arkavo  # required on Homebrew 5.2+
brew install arkavo

Raspberry Pi 5: Download ARM64 binary from releases. See deployment guide for setup. First run auto-selects an edge model for the device (Pi 5 → Gemma 4 E4B).

Install on Windows

Download the installer from the releases page and run the .exe file.

Launch

# Start an agent (zero config)
arkavo

# Or launch web UI
arkavo ui

That's it. No configuration files, no setup. Agents auto-discover via mDNS and form a mesh.

On first run, Arkavo downloads two local models sized to your device — a small model for fast routing (Gemma 4 E2B) and a larger model for inference (Gemma 4 12B on desktop/workstation; Gemma 4 E4B on a Raspberry Pi 5).

Trusting an agent

To authorize an agent (e.g. from another device), show its identity QR — the agent's DID:key and entitlements:

arkavo agent run --trust   # or simply: arkavo --trust

Coming from OpenClaw?

See the migration guide for a full comparison: what you gain (budget controls, TDF encryption, PII preflight, offline operation), what's different, and step-by-step setup.

Why Arkavo?

• Zero config: Just run arkavo. Auto-naming, auto-routing, auto-discovery.
• Fast: Low-latency agent-to-agent communication (benchmarkable from source — see Building from Source).
• Visual: See live agent communication flows in real-time.

SwarmKit

Declarative multi-agent kits where each role declares its own TDF Attribute Release Policy. The orchestrator constructs role-scoped policies before any data reaches the role — push the trust boundary inward.

# Launch any kit at gateway boot
ARKAVO_SWARMKIT_PATH=examples/code-review-kit/code-review-kit.swarmkit.yaml arkavo

Four shipped kits: campaign-kit, code-review-kit, vrm-production-kit, compliance-kit. Full guide: docs/SWARMKIT.md. To validate a kit manifest from source, see Building from Source.

Features

• SwarmKit - Declarative multi-agent kits with per-role TDF attribute-release policies. Four shipped examples covering marketing, code review, creative, and regulated domains. See docs/SWARMKIT.md.
• Multi-provider routing (OpenAI, Anthropic, Gemini, Kimi, DeepSeek, local models)
• Local edge models via llama.cpp - Gemma 4 (E2B/E4B/12B) by default; Ministral 3B/8B (with vision) also supported
• Cost-aware model selection (real per-token estimates on full macOS/Linux builds; the musl-slim and Windows binaries use an approximate estimator)
• iOS simulator automation (macOS only)
• Security scanning (Semgrep, OSV, SBOM)

Usage Examples

Chat

# Use any provider with API key
GEMINI_API_KEY=your-key arkavo chat --prompt "Hello"
DEEPSEEK_API_KEY=your-key arkavo chat --prompt "Explain Rust"

Context Control Demo

The Autonomous Refactor demo demonstrates Active Context Management. It simulates a large-scale "breaking change" refactor that generates extensive compiler output, showing how the Context Ledger maintains a small active window while preserving data access.

cd examples/autonomous_refactor
./run_demo.sh

Custom Agent Config (Optional)

arkavo agent init my-agent  # Creates AGENTS.md template
# Edit AGENTS.md to set model, capabilities, API keys
arkavo  # Runs with your config

Security (Optional)

OpenTDF Integration: Fine-grained access control for MCP tools via OpenTDF. Set OPENTDF_BASE_URL, OIDC_ISSUER, and AUD environment variables.

Coding Agent Toolset

The agent uses these MCP tools in-process during chat and task — there's no separate server to start. Tools that shell out to an external binary register only when that binary is on PATH (noted below).

Code Search & Intelligence

• codegrep_search: Fast repository-wide code search with ripgrep
• struct_find_replace: Language-aware structural search and replace with Comby
• syntax_tree: AST parsing for syntax-aware code analysis with tree-sitter

Security & Quality (require the named binary on PATH)

• sec_semgrep: SAST scanning with Semgrep
• deps_osv: Dependency vulnerability scanning with OSV-Scanner
• sbom_syft: SBOM generation with Syft

Test & Automation

• browser_cdp: Chrome DevTools Protocol automation via chromiumoxide
• test_run: Multi-language test runner (pytest, jest, go test, cargo test, xcodebuild)

Ephemeral Workspaces (requires Docker/Podman)

• workspace_container: Container-based isolated execution with resource quotas

These nine tools are the ones the running agent can call (the binary also registers git, GitHub, web-search, shell, and TDF tools — see the full reference below).

Benchmark harness

SWE-bench evaluation lives in the separate arkavo-mcp-bench crate, run from source — it is not a registered agent tool (it depends on the orchestration engine, which would form a dependency cycle if exposed through the tool registry).

See docs/coding-agent-toolset.md for complete tool documentation.

Security Status

For offensive-security reviewers: here's what is real today and what is still on the roadmap.

Shipping now

Roadmap / not yet landed

This split is intentional: encryption, access control, and identity are shipping now; hardware-bound trust roots are being built in the open.

Platform Support

mDNS discovery uses pure Rust implementation (mdns-sd crate) with no system dependencies

Note: iOS simulator automation and testing capabilities are only available on macOS.

Building from Source

Prerequisites

Install required build tools:

# macOS
brew install cmake ccache

# Linux (Debian/Ubuntu)
sudo apt install cmake ccache build-essential

# Linux (Fedora)
sudo dnf install cmake ccache gcc-c++

Setup llama.cpp

Clone the llama.cpp dependency (not tracked in git):

git clone https://github.com/ggerganov/llama.cpp vendor/llama.cpp
cd vendor/llama.cpp
git checkout f280b26983ad0fdb705a0d9ebf0503e76f2899b0  # b10615
cd ../..

Build

cargo build

The default build includes mDNS discovery using a pure Rust implementation (mdns-sd crate) that doesn't require system libraries like Avahi or Bonjour. This provides true portability across all platforms.

Development

These commands run against the source tree (not the installed binary):

# Measure agent-to-agent latency
cargo bench -p arkavo-protocol --bench a2a_latency

# Validate a SwarmKit manifest
cargo run -p arkavo-swarmkit --example validate_kit -- \
  examples/compliance-kit/compliance-kit.swarmkit.yaml