CONCEPT PREVIEW — simulated fleet data · no public product access yetCOMING SOON
FLEETNEXUS CONCEPT · COMING SOON

ONE DECK.
THE WHOLE FLEET.

Command a legion of autonomous agents from a single console — every loop, every fault, every token, one source of truth. Issue one order, route it through each agent's own gateway, watch it execute. Nothing claims all-green while a backend is down.

A concept for teams running 5–50 always-on agents.

auto-discovers agents on diskzero-inbound remoteoperator-token gated + audited
Fleet · Command Center LIVE · 5 AGENTS
NEOgateway · model · mem3 loops
ORIONgateway · model · mem2 loops
ATLASoauth token · 36h4 loops
VEGAgateway hung1 loop
LYRAgateway · model · mem2 loops
Fleet health94%
Weekly reserve61%
TOKEN BURN14.2k /min
probeneo · gateway 200 · model live · mem fresh
COMMAND CENTER · TELEMETRYsample data
Status
5
agents live
Loops live
42
hermes crons
Watchdogs
7
gated
Token burners
3
flagged
Weekly reserve
$1.2k
headroom
Vault risk
2
on-disk keys
THE DECK

IT'S NOT FIVE TOOLS. IT'S ONE DECK.

Every navigation group is a station on the same console, drinking from the same shared health model — and in every one, the fix sits next to the evidence, not three tabs away.

GROUP 01COMMANDGROUP 02FLEETGROUP 03MEMORYGROUP 04ECONOMYGROUP 05OPSSHARED STATE01 SOURCEONE HEALTHMODELPROBE OUTPUT · SYNCED

Every station reads the same probe output. The snapshot, the CLI table, and the live fault map are the same function called twice — they cannot disagree.

GROUP 01COMMANDRoster, the command console, slash-goal authoring, and the standards-enforcing loop composer.
GROUP 02FLEETAgents drill-in (logs, checks, config, identity) and Faults with deterministic fixes beside the evidence.
GROUP 03MEMORYGBrain explorers, topology, and recovery readiness — backups surface as faults, not surprises.
GROUP 04ECONOMYPer-agent cost, context composition, and waste leaders — intended to be collector-backed in production; sample values in this concept.
GROUP 05OPSAlerts, governance + scaffolds, the Bitwarden credentials matrix, live log feed + fleet timeline.
COMMAND-CAPABLE, SAFELY

You give the order. Every agent's own gateway carries it out.

Type one prompt. Broadcast it across the fleet or one-shot a single agent. The order streams through each agent's native gateway — not a hand-edited backdoor, not a side channel.

Operator-token gated — a second factor on top of Cloudflare Access.
Append-only audit — every mutation logged to state/operation_runs.jsonl.
No hand-edited state — every action runs through an existing adapter: hermes CLI, monitoring, docker, launchctl.

“The HUD never reaches behind the curtain. It pulls the same levers you would — just faster, and on the whole fleet at once.”

Command ConsoleBROADCAST
hermes -z "rotate gateway token; verify model live"
broadcast → 5 agents · gateway env resolved
FAULT → FIX

EVERY FAULT CARRIES ITS OWN FIX.

A fault isn't a notification you chase in another tab. The deterministic fix sits right next to the evidence that justifies it — one click, executed through the agent's own adapter, and audited on the way out.

Fault → Fixthe fix lives next to the evidence
FAULT
vega · gateway hung
model probe: live ping, no completion
SINGLE SOURCE OF TRUTH

NOTHING CAN CLAIM ALL-GREEN WHILE A BACKEND IS DOWN.

Most dashboards report. This one cannot be wrong about whether your fleet is up. The collector doesn't compute health — it reuses the exact probes the live Faults view reads. And the model probe is fail-closed: it sends a real prompt through the agent's model path and requires a completion back. A green light that's six hours old is not green.

2 VIEWS · 1 CODE PATH · 0 WAYS TO DISAGREE.

Generic status page · cached
100% — ALL SYSTEMS GO
gbrain-dboperational
last checked6h 14m ago
▲ reporting a port it pinged once, not a model that answered
F.L.E.E.T · live boardPROBES
CRIT — 1 AGENT DOWN
gbrain-dbunreachable
model probelive ping: no completion
● fail-closed · cannot read OK while down · board agrees with the snapshot
One code pathfleet_probes.py
◈ fleet_probes.py — the only ground truth
├─▶COLLECTOR  state/fleet_status.json (every 180s)
├─▶LIVE FAULTS  the HUD's fault map
├─▶MCP SERVER  check / fix for the agents themselves
└─▶CLI  bin/fleet · same snapshot, rendered for a human

OK means a model answered. Not that a port did.

AGENT FINOPS

A SUB-HOURLY LLM LOOP WITHOUT A PRECHECK GATE IS REJECTED AT CREATION.

This concept models a runaway loop consuming ~25% of a sample fleet bill: a loop firing every 15 minutes with no no-LLM precheck gate — 96 cold model calls a day, zero of them needed. Fleet doesn't warn you after the burn. It refuses to let you create it.

The woundtoken-economics

~25% of a sample fleet bill, burned by one runaway loop:

25%
scout-03$612/wk · 71% redundant context
relay-11$408/wk · */15 llm-loop · no gate

The production design reads collector datasets; this concept uses simulated sample values.

Loop composercron-token-standard
BLOCKED*/15 * * * * llm-loopno precheck gate ✗ standard violation
cadence < 1h → gate requiredllm-on-tick → gate requiredcost ceiling per loop → enforced
Projected model spend96 cold calls / day
HOW THE PIECES FIT

TIMERS IN. ONE MODEL OUT. EVERY SURFACE AGREES.

launchd timers run the collector and the economics jobs on a cadence; they write one shared state. The CLI, the HUD, and the MCP server all read it — so no two surfaces can ever tell a different story.

Data flowlaunchd → state → surfaces
fleet-monitor · 180sfleet_status.py → state/fleet_status.json + _history.jsonl
token-economics · 5mtoken_economics.py → token_economics.json + _history.jsonl
models-registry · 6hmodels_registry.py → models.json
fallback-monitorhermes_fallback_monitor.py → hermes_fallback_monitor.json
state/ (shared)read by  bin/fleet · HUD · fleet_mcp.py — read-only; fixes are separate, authorized, audited steps
REMOTE ACCESS

A COMMAND DECK WITH NO DOOR TO KICK IN.

Cloudflare Tunnel + Access identity at the edge means zero inbound ports — nothing to scan, nothing exposed. The architecture is the guarantee: you don't have to trust the vendor — you can read the probe.

Inbound ports0Zero-inbound by design. Nothing to scan, nothing exposed at the edge.
IdentityCloudflare AccessSSO identity verified at the edge before a request ever reaches the deck.
Per mutationOperator tokenA second factor on every state-changing action — gated and audited.
FROM YOUR TERMINAL

THE WHOLE FLEET, ONE COMMAND.

  bin/fleet — fleet health table
$ bin/fleet
FLEET · 5 agents · 1 CRIT · sync 2s ago
neo   ● ok   gateway 200 · model live · mem fresh  3 loops
orion ● ok   gateway 200 · model live · mem fresh  2 loops
lyra  ● ok   gateway 200 · model live · mem fresh  2 loops
atlas ◐ warn oauth token expiring 36h      4 loops
vega  ▲ crit gateway hung — healer restarting  1 loop
$ bin/fleet --agent vega --json

FLEETNEXUS IS COMING SOON.
EXPLORE THE COMMAND DECK.

This is a working product concept, not yet a commercial service. Explore the vision now; access will open later.

Coming SoonReplay the Tour
$ bin/fleet # human health table for the whole fleet
$ bin/fleet --json # the machine snapshot
$ cd hud && npm run dev # the deck → localhost:4000

One collector · live probes · MCP server · CLI · browser HUD — one shared health model.