Skip to content
- Workspace setup, CI, CLI shell with
onboard, agent, status commands
- Core domain types and traits:
Provider, MemoryStore, Tool, Channel
- OpenAI-compatible provider, SQLite memory,
read_file and shell tools
- Agent loop hardening (max iterations, timeouts, event logging)
- TOML config, env overrides, secret redaction, security defaults
- Gateway HTTP server (Axum) with pairing auth, rate limiting, CORS
- WASM plugin sandbox with integrity verification
- Channel integrations (Telegram, Discord, Slack)
- FFI bindings (Swift, Kotlin, Python via UniFFI; Node.js via napi-rs)
- 35+ LLM provider support via OpenAI-compatible interface
- Autonomy levels, OTP approval, audit trails
- Hardware discovery, cron scheduling, skills/SOP engine
- Workspace consolidated from 46 to 16 crates
- Encrypted SQLite with SQLCipher
- Plugin security hardening (path traversal fix, semver, debouncing, file locking)
- Replaced wasmtime with wasmi as default WASM runtime
- Build variant tooling (default, server, minimal)
- 1,400+ tests passing, 0 clippy warnings
- Provider tool definitions (
ToolDefinition, ToolUseRequest, ToolResultMessage)
- Structured tool dispatch in agent loop with text-based fallback
- Conversation message history with
Vec<ConversationMessage>
- Streaming tool use with
ToolCallDelta and SSE parsing
- JSON Schema validation and
agentzero tools list/info/schema CLI commands
- All 50+ tools implement
input_schema()
- Streaming agent loop —
Agent::respond_streaming() with StreamSink / StreamChunk
- Runtime streaming channel —
run_agent_streaming() returning receiver + join handle
- CLI
--stream flag — agentzero agent --stream -m "hello"
- System prompt support —
system_prompt in AgentConfig, wired through all providers
- Gateway agent wiring — Real agent calls on
/api/chat, /v1/chat/completions, /ws/chat
- SSE streaming — OpenAI-compatible SSE on
/v1/chat/completions?stream=true
- WebSocket streaming — Bidirectional streaming on
/ws/chat
- MCP connection caching —
McpSession with cached subprocess connections and tool schemas
- FFI Node.js parity —
register_tool(), send_message_async(), registered_tool_names()
- Conversation branching and forking
- Multi-modal input (image, audio) across all providers
- Plugin registry and marketplace
- Enhanced RAG with vector search
- iOS XCFramework packaging for Swift FFI
- Android AAR packaging for Kotlin FFI
- Agent-to-agent collaboration protocols
- Cost tracking dashboard
- Distributed agent coordination
- Self-hosted model fine-tuning integration
- Enterprise audit and compliance features
- Add one capability per PR
- Every feature needs: tests, docs, and one explicit non-goal
- All tools must implement
input_schema() for structured tool-use compatibility