LLMVM v2

A next-generation LLM-native agent VM that combines:

LLMVM's philosophy — the LLM is a programmer in a persistent Python REPL with LLM-native opcodes (llm_call, llm_bind, llm_list_bind, coerce, guard, result). Interleaved natural language and code via <helpers> blocks, executed in continuation-passing style.
Claude Code's production substrate — unified Tool interface with capability sets, layered permission system with AST-based plan-mode enforcement, MCP integration, skills-as-data, sub-agent swarms with mailbox IPC, durable transcripts with compact boundaries, worktree isolation.

Quick start

pip install -e .[all]
export ANTHROPIC_API_KEY=...
llmvm repl

See docs/architecture.md or the source under src/llmvm2/:

core/ — continuation loop, transcript, session, content/message/event types
providers/ — Anthropic, OpenAI, Gemini, Bedrock with stop-token normalization
cell/ — sandboxed Python subprocess that executes <helpers> blocks, meta-primitives
tools/ — fs/shell/web/git/browser/etc. with declared capabilities
permissions/ — capability-set engine, modes, AST pre-check
skills/ — markdown + YAML frontmatter, compatible with SKILL.md
memory/ — memdir with MEMORY.md index and post-turn auto-extraction
mcp/ — MCP client (stdio/SSE/HTTP/WS) + deferred tool search
swarm/ — sub-agents, file-based mailboxes, worktree isolation
compile/ — thread compilation with guard-based JIT respecialization
gateway/ — WebSocket JSON (no jsonpickle)
cli/ — REPL and one-shot pipe modes

Active development. See the plan in ~/.cursor/plans/llmvm_v2_architecture_*.plan.md.