Q
Which clients work?
Any MCP-compatible client: Claude Code, Cursor, Codex CLI, Windsurf, Gemini CLI, OpenCode. The protocol is identical; only the config location differs per client.
Docs · MCP Setup (Claude Code, Cursor, …)
MCP Setup (Claude Code, Cursor, …).
One memory, every harness.
Direct answer
Anvaya's MCP server exposes your local graph to any MCP client over stdio: 14 tools, budgeted queries (default 4096 tokens), zero accounts. Five lines of .mcp.json in your project root, restart the client, and Claude Code, Cursor, or Codex CLI reads and writes the same memory Anvaya CLI uses natively.
MCP Setup (Claude Code, Cursor, …)
Steps
011 · Start the daemonMind runs as a Unix-socket daemon (auto-spawn on first client connect). Verify with anv mind stats.
022 · Add .mcp.jsonFive lines in project root naming Anvaya's MCP server over stdio. Same file shape every MCP client expects.
033 · Restart the clientReload Claude Code / Cursor / Codex so it picks up the server. Expect 14 tools: mind_query, mind_search, mind_get_node, mind_record_change, mind_feedback, mind_stats, session start/end, mind_get_threads, mind_delta, mind_proactive, mind_intent, mind_retro_promote, mind_federate.
044 · Query with budgetmind_query takes natural language plus token_budget (4096 default) and returns PackedContext, critical nodes, patterns, fixes, concepts.
055 · Record outcomesAgents should call mind_record_change on decisions and fixes so future sessions (any client) inherit the lesson.
Questions
Asked About MCP Setup (Claude Code, Cursor, …).
Q
Latency vs native?
MCP ~10ms per call vs Direct <0.5ms / IPC ~2ms inside Anvaya CLI. Irrelevant next to model latency; the bridge tax is noise.
Q
Does MCP send code anywhere?
No, stdio to your local daemon. Cloud exposure depends only on which LLM provider your agent itself uses.
Run Agents That Fit On Your Laptop.
25.6 MB median RSS. 25 agents ran in parallel on a Core 2 Duo with 4 GB RAM. Hundreds on your machine. Zero cloud required on the Ollama path.
Requires Rust/cargo to build from source. Linux and macOS today, Windows not yet supported. Pre-1.0, public beta. Pricing TBD.