VisvoAI↗
Sep 2025 — PresentOpen-source Python packages for building AI agents, plus a terminal coding agent and a multi-tenant agent platform built on them. Both extend the packages through hooks and dependency injection, with no forked code.
PythonLangGraphLangChainFastAPINext.jsTextualMCPRAGWeaviateRedisPostgreSQL
Open-source packages (visvoai-ai, visvoai-core)
- visvoai-ai: a provider layer over LangChain chat models with a live model registry (pricing, context window, capabilities), cost metering and one set of reasoning levels across Anthropic, Gemini and OpenAI-compatible APIs.
- visvoai-core: a LangGraph agent runtime with override hooks for nodes, routing, state, checkpointing and persistence, so products extend it without forking, plus a step cap that forces a clean final answer.
- A single tool contract for typed functions, class-based tools and LangChain tools, with self-registration, validation at import and pluggable call logging.
- Hand-built BM25 + cosine tool search, generalized from the platform into the core package.
- CI enforces the boundary between the public packages and the private platform, so private code or names can't reach a public release.
Terminal coding agent (visvoai-cli)
- A terminal coding agent built on the two packages, unmodified, and the reference for how far they go.
- Shell commands are sorted into read or write: reads run inside an OS sandbox (macOS sandbox-exec, Linux bwrap) that can't write to disk, and every write asks for approval first.
- Subagents run in parallel with their own context and live logs, alongside skills and MCP servers that stay off until you trust them.
- Every turn saves a snapshot, so /rewind restores both the files and the conversation to any earlier point.
Platform
- Agent harness and loop on LangGraph: a custom tool-execution node with parallel batching, repeat-call blocking, schema-guided recovery from bad arguments, step caps and context compaction.
- Dynamic tool binding so agents stop loading every tool, which had used ~57% of each prompt: per-step retrieval over Gemini-embedded MCP tools, and tool descriptions rewritten to lift recall@8 from 91% to 97% on a 211-case benchmark.
- A tool system of ~40 tools with roles and per-tool permission checks before execution, plus human-in-the-loop approvals (approve, edit, approve-all, reject) that pause and resume the agent.
- Custom agents users define at runtime, 17 on-demand skills, and subagent calling with isolated context, depth caps and parallel runs that stream live to the parent's UI.
Show 4 moreShow less
- Versioned artifacts (docs, code, slides, sheets, charts) that agents and users co-edit with conflict checks, SQL over spreadsheets in DuckDB, and a network-isolated sandbox where agents run code.
- RAG over a per-user document Library: PDF/DOCX/PPTX parsing, heading-aware chunking, Gemini embeddings and hybrid BM25 + vector search in Weaviate with citations, plus long-term chat memory.
- An MCP client and OAuth connectors written from scratch (Google, Microsoft, Slack, Notion, HubSpot) that sync server tools by content hash, repair malformed LLM arguments and move large results to files.
- LLM cost and token tracking with a trace viewer, automatic audit logs, encrypted per-tenant API keys, conversation branching, and stream resume on page reload over Redis Streams.
