v0.2 · macOS · Linux · WindowsApache-2.0 core · BYOK

One layer underevery coding agent.

One layer that works across Claude Code, Codex, Gemini and Aider — and gives them what none has alone: unified memory of every session, a signed plugin layer, a mobile companion, and an audit trail no agent can rewrite.

Runs your agents side-by-side in a terminal grid — keys in your OS keychain, no proxy, Apache-2.0 core. What we never collect →

Download for macOS ⌘D GitHub See Council Mode →

~/furx▸ Memory Hub · cross-CLI

claude-A · pane 01

$ remember "auth uses bcrypt 12-round salt, rotate keys quarterly"

→ memory.write ok · id mem_3f2a · indexed · 18ms

// 4 seconds later, in another pane…

codex · pane 02

$ recall "auth"

1. mem_3f2a · claude-A · 4s ago

"auth uses bcrypt 12-round salt,

rotate keys quarterly"

cross-CLI memory · SQLite FTS5 · local

memory_hub · 1 write · 1 read · 23ms total~/.furx/memory.db · ∞ retention · 0 bytes leaked

1 / 3 · auto-rotate every 7s

hover to pause

"Every agent forgets when the session ends. Furx is the memory, the rules, and the remote — for all of them."

— the shared substrate, not another terminal

What every agent is missing.

— the four pillars

i · across any CLI

Unified memory

Every session your agents run — searched, recalled, deduped — in one local index. Claude Code, Codex or Gemini; the memory is shared. “Don't repeat what already failed” becomes a query, not a hope.

ii · write once

Unified plugins

One signed plugin layer (Ed25519) that runs in any agent — not a fork per CLI. Skills, tools and gotcha packs you author once and trust everywhere, with a verifiable signature, not a copy-pasted script.

iii · off the desk

Mobile companion

Pair your phone with a QR code and drive live sessions from anywhere on your LAN or Tailscale — read a pane, approve a step, fire a prompt. The agents keep running; you just step away from the desk.

iv · the agent can't rewrite it

Governance

An append-only audit log (local SQLite, enforced by DDL trigger) of every action, plus policy and roles. History the agent — or you — cannot quietly edit. The accountability a room full of autonomous agents actually needs.

Underneath it all: your agents run side-by-side in a terminal grid, and Council Mode (⌘J) fans one prompt across up to six models when you want a second opinion. The pillars are what you can't get anywhere else.

Council Templates by phase.

— Free + Pro

The five built-in presets pick voices by provider type (OpenRouter Quick / Free Tiers / Paid / Local / Mix). Templates pick voices by workflow phase— so the Council you fire matches what you're actually doing.

Template	Default voices	Best for
Planning	Opus · GPT-5 · Qwen-235B · DeepSeek-V3.1	Spec, architecture, high-level decisions. Heavy reasoning, time-tolerant.
Implementation	Sonnet · Codex · Gemini Flash · Aider	Diff & code generation. Speed and code focus.
Review	Opus · GPT-5 · Gemini Pro · Llama-4	PR review, audit. Diverse families for cross-family blind spots.
Debug	Sonnet · Codex · Aider · local qwen-coder	Stack-trace aware. Cloud + local mix to keep velocity.
Refactor	Gemini 2.5 Pro · Opus · GPT-5	Large codebases. Models with >200K-token context window.

All templates ship Free. Override the voices, save your own template, or write a YAML config.

Connect whatever you already have.

— 5 paths

The Furx Connect wizard offers five paths — mix-and-match freely. Furx doesn't manage other people's keys; every call goes straight from your machine to the provider.

OpenRouter Quick Start

1 key · 30s

One key, hundreds of models. Council Mode picks distinct providers from the OpenRouter catalog — Claude, GPT, Gemini, Qwen, DeepSeek, Llama-4 — and dispatches to as many as you wire up.

Free tiers

$0 / forever

Separate free accounts: Cerebras, Groq, Mistral, SambaNova, Gemini AI Studio, OpenRouter free. Daily quotas vary per provider; Furx orchestrates whatever you connect.

Paid APIs direct

your spend

Anthropic, OpenAI, Gemini paid, Cerebras paid, or any OpenAI-compatible endpoint. Your billing, your rate limits, your usage caps. Furx never proxies.

Local inference

$0 · privacy

Auto-detects Ollama (127.0.0.1:11434), LM Studio, llama.cpp, vLLM. Council with local-only models: inference never leaves your machine.

Proxy / LiteLLM

org governance

Paste your org proxy URL. Spend tracking, allow-listing, audit — Furx doesn't care which gateway you use as long as it speaks OpenAI-compatible JSON.

Mix-and-match

Council preset

Combine all five. Council dispatches across whichever providers are healthy. Graceful degradation if one voice rate-limits, hard timeout 40s.

What we never charge for.

— in stone

BYOK means your provider spend is yours. Pro pays for cloud sync, session replay, encrypted backups and cost analytics — services we have to host. Everything below is Free, forever, Apache-2.0.

✓Provider keys (OS Keychain only — Furx can't read them after you store)
✓Prompts you send (machine → provider, never via us)
✓Responses you receive (same)
✓Number of voices in Council Mode (one or all six — same price: $0)
✓Number of panes (open as many as your screen fits)
✓Local audit log in ~/.furx/furx.db
✓Local Memory Hub (SQLite + FTS5 + Knowledge Graph)
✓Skills you write in ~/.furx/skills or ~/.claude/skills
✓Voice transcription (whisper.cpp local, never uploaded)
✓Mobile bridge (point-to-point WS on your LAN or Tailscale)

Pricing, plain.

See detail →

Pass-through means our cost isn't tied to your LLM spend — we charge for the orchestration layer that we host on Cloudflare.

Free

Apache-2.0 · forever

Pro

$12

/mo · 14-day trial

Team

$30

/seat/mo · 5+ seats

Enterprise

$49

/seat or $2.5k perpetual

FAQ

›Do you take a cut of my LLM costs?

Never. Pass-through only. Furx never sees your provider keys — they live in your OS Keychain, sent directly from your machine to the provider. There is no proxy and nothing to mark up.

›What does Pro pay for if not voices?

Cloud sync of your skills, .mcp.json, and Memory Hub backups. Session replay (30 days). Cost Meter Pro with alerts and CSV export. Latency heatmap trends. Encrypted daily backups. Council Mode itself is identical on Free and Pro.

›How many models can the Council dispatch to at once?

Up to six voices per dispatch. The five built-in presets fill all six with OpenRouter Quick Start, and you can swap any voice or save your own preset. Free, Pro, Team — all the same.

›What's the trial?

14 days Pro on first install, no credit card. Auto-reverts to Free if you don't convert. The trial unlocks cloud sync and replay; everything you can do offline keeps working.

›Can I self-host Pro/Team?

Enterprise tier ($49/seat or $2.5k perpetual). Self-hosted notarized build, data residency Argentina or EU, source-code escrow via NCC Group.

›Refund policy?

14-day no-questions refund via Paddle MoR. Paddle handles EU VAT, AR Argentina, US sales tax automatically.

Be there when Furx goes public.

The shared layer for your coding agents — unified memory, signed plugins, mobile companion, governance — is launching open-source, Apache-2.0. Leave your email and we'll tell you the moment the public build is live. No spam, just the launch.

Star on GitHub Sign in to dashboard