Warmplane is the local control plane that multiplexes upstream Model Context Protocol (MCP) servers behind one hot, policyable runtime. Cut token overhead by up to 96% with lazy capability indexes, Multi Round-Trip Requests (MRTR), hybrid vector search, fine-grained RBAC governance, and deterministic execution envelopes.
Before an LLM agent executes a single tool call, it ingests the full schema catalog of every connected MCP server on every turn. As server count grows, prompt tokens balloon rapidly—adding cost, latency, and context distraction.
Reproducible evaluation measured with the cl100k_base tokenizer. Run the evaluation suite in eval/token-efficiency/ to verify on your own workloads.
Suite: github_copilot_only
Discovery 54,715 -> 2,386 (95.6% savings)
5-turn loop 260,005 -> 11,173 (95.7% savings)
10-turn mixed 547,150 -> 22,895 (95.8% savings)
Control: public_filesystem_only
Discovery 2,552 -> 1,066 (58.2% savings)
5-turn loop 12,760 -> 5,349 (58.1% savings)
10-turn mixed 25,520 -> 10,669 (58.2% savings)
Everything ships in a single zero-dependency Rust binary compliant with Microsoft Pragmatic Rust guidelines.
io.modelcontextprotocol/tasks capability: asynchronous long-running tool execution (202 Accepted + resultType: "task"), MRTR inputRequests delegation, cooperative cancellation, and persistent atomic task registry.
idk_<sha256>), X-Warmplane-Deduplicated caching, replay count tracking, WORM audit trail linking, and in-flight cancellation.
profiles), static/JWT token auth, per-role policy overrides, dynamic catalog/search pruning, and automated secret redaction.
EmbeddedWarmplane and ControlPlaneHandle — zero HTTP/IPC serialization overhead on your Tokio runtime.
Explore the primary RESTful HTTP/v1 endpoints exposed by Warmplane daemon.
All interfaces share identical policy rules, session warmers, idempotency layers, and execution envelopes.
/v1/ API exposing capability indices, SEP-2663 tasks, hybrid search, SSE streams, and operation cancellation.
EmbeddedWarmplane, ControlPlaneHandle) with zero HTTP/IPC overhead.
Configure your upstreams, validate your profile, and start serving warm sessions.
{
"port": 9090,
"toolTimeoutMs": 15000,
"capabilityAliases": {
"sqlite.read_query": "db.query"
},
"policy": {
"allow": ["db.*", "fs.*"],
"deny": ["fs.secret"],
"redactKeys": ["token", "api_key"]
},
"mcpServers": {
"sqlite": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-sqlite", "./db"]
}
}
}
# Install binary from source or crates.io $ cargo install warmplane # Interactively add servers, or import from Claude Desktop / Cursor $ warmplane server add $ warmplane config import # Start the local control plane daemon $ warmplane daemon --config mcp_servers.json # Perform vector/keyword hybrid tool search $ warmplane search-capabilities "query database" --limit 5 # Execute tool through control plane $ warmplane call-capability db.query \ --params '{"query":"SELECT 1"}' \ --idempotency-key op-42-run-1
Start cutting AI agent token overhead while adding enterprise policy control.