Local AI Workflow Engine
A multi-agent orchestration engine for your desktop. Runs high-quality open models locally, indexes your codebase with a hybrid polyretrieve RAG pipeline, and exposes everything over MCP — so your phone can dispatch tasks, and your Artel workspace stays in sync.
Flow Console is a local-first agentic engineering tool that runs a full multi-agent orchestration layer on your machine. A dedicated Orchestrator agent decomposes complex engineering tasks into a directed execution plan, spawns specialist worker agents — each scoped to a precise objective — and coordinates their output into verified results. Every step runs against your actual codebase, with no round-trips to a cloud planner.
The engine runs high-quality open models entirely on-device via llama.cpp. RAG is handled by a local polyretrieve pipeline — hybrid dense semantic search combined with sparse keyword recall — keeping your code and docs on your machine at all times. Cloud LLMs are an opt-in acceleration path, not a dependency.
Flow Console runs its entire inference and retrieval stack locally. The embedding model is bundled and indexes your project files on first open, then updates incrementally as you work. Agent reasoning, tool calls, file edits, and retrieval all happen on-device — no telemetry, no cloud dependency. Cloud LLMs become an optional acceleration tier you control, not infrastructure you depend on.
Flow Console exposes its full agentic surface over the Elenchus dispatch protocol. From the Elenchus mobile app, you can start tasks, watch the orchestrator's live thinking and tool execution stream, stop a running agent mid-task, or queue follow-up operations — all without sitting at your desk. Sessions persist across remote interactions, so a task started from anywhere runs to completion on your machine.
The installer ships with the embedding model built in — local RAG works from first launch, no download required. Chat models install separately as tiered packs inside the app: the Standard pack (Gemma 4B + Qwen 2.5 Coder 7B, ~11 GB) is free for all users. Cloud-only users can skip the download entirely. Premium and Ultra packs with larger, higher-capability models are coming for Pro and Unlimited subscribers.
~730 MB installer · embedding model included · chat models (Standard pack: Gemma + Qwen, ~11 GB) download inside the app · no API keys required
Flow Console runs as a persistent MCP server on the local machine, registering its agentic tools as callable resources. Connected clients — Elenchus for dispatch control, Artel for workspace context — invoke tools over a single, consistent protocol surface. Computation stays on the user's machine; control travels anywhere.
The architecture separates orchestration, execution, and context cleanly. The Orchestrator holds the task graph. Worker agents hold execution state. The MCP server exposes the interface. Artel reads and writes context without breaking the locality guarantee — five live tools, ACL-enforced server-side, working live against your workspace.
Flow Console doesn't ship in isolation. It plugs into the rest of the Metantel stack through a shared protocol surface — context flows in and out without leaving the family.
Elenchus is the mobile command centre — start agentic tasks, watch the live execution stream, stop or queue ops, all from your phone.
Open Elenchus →Five live MCP tools let Flow Console agents read project context, log decisions, update tasks, and write back to Artel docs — all ACL-enforced.
Open Artel →