Warmplane

The local control plane that keeps MCP sessions warm.
v0.14.0 — Changelog · User Guide · Performance · Whitepaper · OpenAPI

Warmplane runs multiple upstream MCP servers behind one local process, keeps those sessions persistent, and exposes a compact, policy-aware surface for tools, resources, and prompts — accessible via HTTP, CLI, and MCP-native clients.

Quick Start

1. Build

cargo install --path .

# Optional: local ONNX vector embeddings (FastEmbed)
cargo install --path . --features semantic-search

2. Configure Upstream Servers

Manage servers interactively or import from existing tools:

# Interactive setup wizard
warmplane server add

# Or add non-interactively
warmplane server add filesystem --command npx --arg "-y" --arg "@modelcontextprotocol/server-filesystem" --arg "/tmp"
warmplane server add context7 --url "https://mcp.context7.ai/sse" --bearer-env "CONTEXT7_API_KEY"

# Or import directly from Claude Desktop / Cursor
warmplane config import

Or manually create mcp_servers.json:

{
  "port": 9090,
  "toolTimeoutMs": 15000,
  "capabilityAliases": { "sqlite.read_query": "db.query" },
  "resourceAliases":  { "filesystem.file:///tmp/readme.txt": "fs.readme" },
  "promptAliases":    { "github.code_review": "prompt.code-review" },
  "policy": {
    "allow": ["db.*", "fs.*", "prompt.*"],
    "deny":  ["fs.secret"],
    "redactKeys": ["token", "api_key", "password"]
  },
  "mcpServers": {
    "sqlite":     { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-sqlite", "./test.db"] },
    "filesystem": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"] }
  }
}

3. Validate, then start

warmplane server list
warmplane validate-config --config mcp_servers.json
warmplane daemon --config mcp_servers.json

Run Modes

All three modes share the same backend state, aliases, policy checks, and timeout behaviour.

HTTP Daemon

warmplane daemon --config mcp_servers.json
# Serves /v1/... on the configured port (default 9090)

Key endpoints:

MCP Server (stdio)

warmplane mcp-server --config mcp_servers.json

Point any MCP-native client at this process. It exposes lightweight facade tools (capabilities_list, capability_search, capability_describe, capability_call, capabilities_batch_call, resource_read, prompt_get, …) alongside native resources/* and prompts/* methods.

Claude Desktop / Cursor config:

{
  "mcpServers": {
    "warmplane": {
      "command": "warmplane",
      "args": ["mcp-server", "--config", "mcp_servers.json"]
    }
  }
}

CLI Configuration & Operations

# Interactive server setup wizard
warmplane server add

# Import settings from Claude Desktop or Cursor
warmplane config import

# Hot-reload in-memory workers from mcp_servers.json without restarting
warmplane reload

# Inspect & test servers
warmplane server list
warmplane server test github

# Ecosystem config import (Claude Desktop, Cursor, Zed)
warmplane config import

# Aliases and Policies
warmplane config alias set tool git-commit github.create_commit
warmplane config policy allow "github.*" "fetch.*"

# Capability and Execution CLI
warmplane list-capabilities
warmplane search-capabilities "triage logs" --limit 5
warmplane describe-capability db.query

warmplane call-capability db.query \
  --params '{"query":"SELECT 1"}' \
  --request-id req-101 --actor-id user-7 \
  --idempotency-key op-20-run-1

warmplane read-resource fs.readme
warmplane get-prompt prompt.code-review --arguments '{"code":"fn main() {}"}'

warmplane list-catalog-events --after evt_3
warmplane cancel-operation req-101

Performance Highlights

Warmplane is engineered with pure Rust zero-cost abstractions, keeping agent loops snappy:

• 50.4 ns ETag Cache Validation (If-None-Match $\rightarrow$ 304 Not Modified)
• 159.8 ns Idempotent Cache-Hit Deduplication
• 1.58 µs SHA-256 Incremental Catalog Version Hashing ($N=10$)
• 15.9 µs Filtered Hybrid Capability Search ($N=50$)
• 371 µs In-Memory Zero-Allocation Lexical Tag Search across 1,000 Tools

👉 See the complete benchmarks and profiling methodology in docs/PERFORMANCE.md.

Feature Overview

Changelog

v0.14.0 — Agent Enrichment Suite: Facade Search, Context Distillation & Multi-Step Batching

• MCP Facade Hybrid Search: Exposed capability_search tool over MCP stdio facade with keyword, tag, server ID, and execution mode filters.
• Context Distillation & Truncation: Added _jsonpath, _limit_lines, and _truncate_bytes modifiers to /v1/tools/call and capability_call to protect LLM context windows from oversized outputs.
• Multi-Step Chained Batch Calls: Added POST /v1/tools/batch_call and capabilities_batch_call for single-roundtrip dependent tool executions with $step.field reference interpolation.

v0.12.0 — Cryptographic WORM Audit Log & SIEM Streaming

• Append-only linear SHA-256 hash chaining over tool execution and HITL decision events.
• Audit verification and export endpoints (/v1/audit/...), Splunk HEC and generic Webhook ingestion.

v0.11.0 — Control Deck Web UI & Dynamic Upstream Hot-Reloading

• Control Deck Web UI: Embedded zero-dependency web dashboard at /ui and / with live telemetry, server manager, interactive tool playground, policy/redaction manager, and alias registry.
• Zero-Downtime Dynamic Upstream Mounting: Dynamically mount and unmount stdio, HTTP, and OAuth2 upstream workers via REST/UI/CLI without restarting the daemon process.
• Dynamic Catalogs & ETags: Concurrency-safe state management with automatic SHA256 ETag recomputation and SSE resource notifications on server changes.
• Explicit Config Hot-Reload: Added warmplane reload CLI command and POST /v1/config/reload endpoint to cleanly reconcile in-memory workers against manual edits to mcp_servers.json.

v0.10.0 — CLI Configuration Management & Ecosystem Importers

• Added warmplane server (add, remove, list, get, test) commands with interactive inquire wizards and flag-driven headless automation.
• Added warmplane config (init, show, import, alias, policy) for safe, transactional configuration mutations.
• Added auto-discovery and import from Claude Desktop, Cursor, and Zed settings.
• Added atomic configuration file writes (fs::rename) preventing JSON corruption.

v0.9.0 — MCP 2026-07-28 Spec Compliance, MRTR & Subscriptions

• Upgraded to official MCP Rust SDK rmcp 3.1.2 and reqwest 0.13.
• Set default protocol version to "2026-07-28" with backward compatibility for "2025-11-25".
• Added Multi Round-Trip Requests (MRTR) support (input_responses and request_state) across HTTP REST, stdio facade, and upstream workers.
• Added cache hints (ttl_ms: 300000, cache_scope: "public") and deterministic alphabetical sorting on catalog listings.
• Added subscriptions_listen tool to stdio facade and /v1/resources/updates SSE stream for real-time notifications.
• Validated RFC 9207 / SEP-2468 iss callback verification for OAuth 2.0 flows.

v0.7.0 — Pragmatic Rust Modernization & Builder Patterns

Full adoption of Microsoft's Pragmatic Rust Guidelines (AGENTS.md). Implemented Builder Pattern (M-INIT-BUILDER) for core state (AppStateBuilder), search filters (SearchFilterBuilder), and request context (RequestContextBuilder). Enhanced error safety (M-PANIC-IS-STOP), structured logging (M-LOG-STRUCTURED), canonical documentation (M-CANONICAL-DOCS), and flexible trait interop (M-IMPL-ASREF).

v0.6.0 — Idempotency, Cancellation & Retry Metadata

Pass Idempotency-Key / X-Idempotency-Key to deduplicate concurrent tool calls. Abort any in-flight request via cancel endpoint or CLI. Every response envelope now includes structured "retry" metadata (classification + state) for orchestrator-aware retry logic.

v0.5.0 — Request Context & Correlation

Structured RequestContext (operation_id, work_item_id, actor_id, grant_id) threaded through all execution envelopes and tracing spans. HTTP header fallback (X-Request-ID, X-Operation-ID, X-Actor-ID, X-Grant-ID).

v0.4.0 — Catalog Versioning & Cache Validation

SHA-256 catalog version. ETag headers on all catalog reads, If-None-Match conditional requests returning 304 Not Modified. GET /v1/catalog/events change feed with cursor-based pagination.

v0.3.0 — Hybrid Search

POST /v1/capabilities/search with BM25 scoring, optional FastEmbed vector embeddings, tag/server-ID filters, and ranked results.

Docs