Run DAGENTX locally

One line. Pulls the CLI binary, installs Ollama if missing, downloads gemma 4 (the default brain), runs the interactive setup wizard.

curl -fsSL https://farnsworth.dagentx.com/install.sh | bash

What it does

Downloads dagentx for your platform → ~/.dagentx/bin/ (symlinked into ~/.local/bin/).
Installs Ollama (skip if already present).
Pulls gemma 4 (~9 GB, default brain) and nomic-embed-text (~280 MB, embeddings for memory recall).
Runs dagentx init — interactive wizard for API keys (Anthropic, OpenAI, xAI, Groq, DeepSeek, Mistral) and additional Ollama models.
Initializes the 7-layer memory store (SQLite + WAL + AES-256-GCM at-rest encryption when a key is set) at ~/.dagentx/dagentx.db.
Writes config to ~/.dagentx/.env (mode 0600).

After install

dagentx http --bind 127.0.0.1:8080   # local web app
dagentx deliberate "your question"   # one-shot CLI
dagentx mcp                          # MCP server (Claude Code, Gemini CLI, …)

Direct binary (no script)

Linux x86_64 — dagentx-x86_64-unknown-linux-gnu
macOS / Windows: build from source (cargo build --release -p dagentx-cli)

Ask the swarm.

A heterogeneous committee proposes, critiques, refines, and votes — server-side and (optionally) right in your browser via WebGPU.

skip memory include browser models

Memory

7 layers, hybrid retrieval, full provenance.

History

Every persisted deliberation, full transcript.

Pick a deliberation to see the full transcript.

Committee

Server-side agents, the judge, and any browser models you've loaded.

Each agent is an LlmProvider impl — Claude, GPT, Gemini, Ollama-local, WebLLM. The committee runs four rounds (propose → critique → refine → vote) under a strategy of your choice. Adaptive picks per-question. The judge is itself an agent — typically the highest-quality one — that scores final answers and writes the consensus.

Judge

Strategies

Browser Models

Run an LLM right in this tab using WebGPU. Loaded models join the committee alongside the server's API agents.

Provider Mode

Share your browser's GPU. Earn DAGNT for verified jobs. When the paid queue is empty your node still works — pulling curiosity-driven self-improvement jobs.

How it works

1
Load a model.
Visit Browser Models and download one of the WebLLM models (Llama 3.1, Phi-3.5, Qwen 2.5, Hermes 3, …). Weights cache in your browser; inference runs entirely on your GPU.
2
Connect your wallet (optional but recommended).
Click Connect wallet top-right. EIP-191 nonce-sign binds your DogeOS address to a provider identity (anti-sybil). With a wallet, you can register a payout address and settle DAGNT on-chain. Without one, you still earn off-ledger units that you can later attach to a wallet.
3
Toggle Share compute.
The page polls /api/compute/jobs/poll every few seconds. Each job is run on your loaded model; the response is submitted with a self-reported quality score. The orchestrator verifies (~1% redundancy spot-check) and credits your ledger.
4
Settle DAGNT.
On the Training tab, register your payout address and press Settle. Default: 1000 units = 1 DAGNT. Reward multiplier = base × quality × scarcity × reputation (capped 1.5×). Slashed nodes can drop below 1× until they recover.

Job kinds the orchestrator dispatches: inference, embedding, eval_rollout, synth_data_gen, curiosity. Browser-class nodes get the first five. Reputation starts at 50/500; each verified job +1 to +1.5, each slash −3. Nodes below −10 are blocked. Quality is your self-report (0–1) plus the agreement check on redundant jobs.

Share compute

Identity

—

Jobs done

Units earned

Status

idle

Live job stream

Recent jobs

Ledger is the canonical record of your earnings. It's persisted server-side and used by the DAGNT settler when you register a payout address.

CROSS-CHAIN COLLECTION

DAGENTX NFTs

Each token lives on DogeOS Chikyu. The image is reconstructed from raw Monad calldata via the DropClaw chunk protocol. Click a card for the full metadata, on-chain explorer links, and chain-flow diagram.

DogeOS · token Monad · image

Connect a wallet, give it a name, and we'll mint it on Chikyu for you.

only mine —

Token ownership lives on DogeOS (contract 0x889F…5c2d). Image bytes are split into chunks and stored as raw Monad calldata via the DropClaw protocol — cheap multi-MB storage, content-addressed. The gallery reads tokenURI + reassembles Monad chunks on demand.

Sniper

Live DEX pair feed → AlphaSniper-7 committee analysis → trading signals. Paper-trade ledger included; real broadcasts stay behind the wallet.

Each pair gets pulled from the configured DEX adapter (env DAGENTX_DEX_API for production; stub by default). The AlphaSniper-7 system prompt drives the committee, and outputs are parsed into a structured signal: BUY/SELL/HOLD/WAIT, entry, exit, stop-loss, risk 1-10, confidence 1-10, timeframe, position size, rationale. Open a paper position from any signal; close it later with the actual exit price. Realized PnL feeds back into the training corpus as a SniperOutcome event.

Manual analyze skip risk score

Recent signals

Watchlist connect wallet to use

intent: auto-buy

Paper positions

Realized PnL

—

Unrealized PnL

—

Open

Closed

Training Network

Every connected node feeds the model. Compute earns DAGNT; useful interactions flow into the corpus. The network is never idle — when paying jobs run dry, nodes pull self-improvement work straight from the curiosity engine.

How it works. Three layers: (1) capture — every chat/deliberation/preference/sniper-outcome lands in a content-addressed corpus, gated by your consent setting; (2) compute — orchestrator dispatches jobs (inference/embedding/eval/synth/lora-step/distillation) to connected nodes, ~1% redundancy spot-checks, slashes liars; (3) settle — verified units convert to DAGNT (1000 units = 1 DAGNT by default). Scarcity, quality, reputation each multiply the reward. Connect wallet → register payout address → press Settle. Until DAGENTX_TOKEN_CONTRACT is set the ledger runs in pure off-chain mode.

Corpus events

—

Opt-in

—

Avg quality

—

Compute queue

—

In flight

—

Units paid

—

DAGNT

loading…

Leaderboard

By kind

Recent events

Export corpus

opt-in only

Settings

Local preferences live in your browser. Server config lives in .env.

Theme

midnight aurora paper

Aurora is greener; Paper is a high-contrast light theme.

Motion

reduce animations

Disables the volumetric WebGL background and most CSS transitions. Reload to apply.

Training consent (default)

opt-in by default ask each time opt-out

Controls whether new chats/deliberations enter the training corpus. Existing entries aren't affected — change them per-event in History.

Server

/

Origin the SPA is served from. Change DAGENTX_BIND on the server to relocate.

WS: connecting… DAGENTX 3.0

Hotkeys

? open help
Esc close modal
/ focus composer
Ctrl+Enter deliberate
g+c Chat
g+m Memory
g+h History
g+b Browser
g+p Provider
g+t Training
g+s Sniper
g+n NFTs

Onboarding

Replay the welcome banner.

Local data

Theme, motion preference, dismissed banners, provider identity.

Documentation

DAGENTX network

Run DAGENTX locally

What it does

After install

Direct binary (no script)

Welcome to DAGENTX

DAGENTX — Quick reference

Hotkeys

Concepts

Ask the swarm.

Winner

Memory

History

Committee

Judge

Strategies

Browser Models

Provider Mode

How it works

Live job stream

Recent jobs

DAGENTX NFTs

Sniper

Recent signals

Watchlist connect wallet to use

Paper positions

Training Network

DAGNT

Leaderboard

By kind

Recent events

Export corpus

Settings

DAGENTX network

Welcome to DAGENTX

Ask the swarm.

Winner

Memory ?

History

Committee ?

Judge

Strategies

Browser Models ?

Provider Mode ?

How it works

Live job stream

Recent jobs

DAGENTX NFTs

Roll your DAGENTX NFT

Sniper ?

Recent signals

Watchlist connect wallet to use

Paper positions

Training Network ?

DAGNT

Leaderboard

By kind

Recent events

Export corpus

Settings

Memory

Committee

Browser Models

Provider Mode

Sniper

Training Network