Agent infrastructure, in the open

Host-agnostic multi-agent runtime

• Turns a single machine into a durable multi-agent node, running Claude Code, Codex, and Maestro backends with provider-native sessions preserved across turns.
• Generalizes Clawgram's orchestration into a reusable core — composable modules and local-first encrypted state behind a single adapter boundary, so new hosts plug into the same runtime.

Embeddable TypeScript agent SDK

• A composable agent SDK that reaches Claude Code-level agentic coding performance on non-Anthropic backends (DeepSeek, Kimi) — evidence that the agent loop, not the model, is the differentiator.
• Sessions, memory compaction, MCP tools, guardrails, and subagent delegation are independently composable behind one streaming interface.

Telegram control layer for Claude Code

• In daily use by practicing lawyers on real case work. Private Mode masks PII with a pre-LLM hook before any prompt reaches the model, then unmasks on the response — confidential documents never leave the device unredacted.
• A multi-tenant orchestration layer maps Telegram forum topics to persistent, isolated Claude Code sessions, with MCP primitives that let sessions exchange context and delegate asynchronously.

Stealth-first browser automation over MCP

• A high-performance browser automation server in Rust exposed only via MCP — a ~5MB single binary alternative to Node-based automation stacks.
• Chrome DevTools Protocol client and browser engine written from scratch (ab-cdp, ab-browser, ab-mcp), with multi-tenant tab ownership so many MCP clients share one browser process safely.

Classical ML implemented from scratch — Gaussian EM, Gaussian processes, HMMs, Fisher LDA, mixture density networks, matrix factorization, PCA, probabilistic graphical models, and restricted Boltzmann machines.