Pre-launch Private waitlist open

Your AI team.
Your VPS.
Your rules.

A self-hosted, multi-model agent platform — governed by shell scripts and cron, not by prompts the model can rationalize past. Bring any cheap model. The system makes it safe, observable, and yours.

agent.bytesmith / run summary COMPLETED

09:14:03conciergeTask dispatched to engineering

09:14:05engineeringModel: deepseek-chat via anthropic-compat

09:14:07engineeringCache hit rate 97%

09:58:44engineeringCompleted 41 turns in 44m 37s

Cost $0.05

Tokens 1.2M

Interrupt ready

Requirements: a Linux VPS, Slack admin rights,
a model API key (any Anthropic-compatible endpoint)

Pixel-art agent pointing — the concierge

concierge listening

Governance

Deterministic interrupt

Any human message in Slack freezes the entire swarm via a flag file — mechanical, not instructed. Born from a real postmortem: agents ignored a "Stop!" message. Now the interrupt is a shell primitive, not a polite instruction the model can rationalize past.

Zero-token governance

Spend caps, watchdogs, and health checks are shell + cron, burning zero tokens. Critical behavior is enforced by the harness. The model never sees governance logic it could talk itself out of.

Least-privilege capabilities

Every integration — Slack, email, GitHub, image generation — is a plug-in behind one contract. Scoped, revocable credentials only. No broad secret is ever within an agent's reach.

Model swappable at runtime

The harness is Claude Code; the inference target is not. Point any agent at any Anthropic-compatible endpoint — DeepSeek, Kimi, or whatever is cheapest this quarter. No code change required.

Real numbers

Cost is a first-class
design constraint.

A concierge dispatch run costs $0.029. A 41-turn engineering run on DeepSeek cost $0.05 — measured, not estimated.

The same workload on Kimi K2.6 ran 20× more expensive. A 97% prompt-cache hit rate is the lever.

These are production numbers from the platform building itself. We dogfood every layer before shipping it.

engineer

$0.029

concierge run cost — dispatches, routes, confirms

$0.05

41-turn engineering run on DeepSeek

97%

prompt-cache hit rate on sustained workloads

~20×

cost reduction vs Kimi K2.6 on equivalent task

POSTMORTEM Incident #001 — autonomous agent ignored "Stop!"

"Stop" doesn't mean anything
to a reasoning model.

During an early autonomous run, a "Stop!" message was sent to the Slack channel. The agents continued. The model processed the instruction as data — weighed it, found reasons to proceed.

The fix wasn't a better prompt. The fix was a flag file checked by the harness before every tool call. Now, a human message touches the filesystem. The agents stop — not because they understand the instruction, but because the harness enforces it mechanically.

        # checked before every tool invocation

        if [ -f /tmp/bytesmith/interrupt.flag ]; then

          echo "INTERRUPTED — human in command"

          exit 1

        fi

reviewer

The shift

Self-hosting AI went mainstream.
Workspaces first. Teams next.

62k GitHub stars in a week for Odysseus settled the argument: people want AI on hardware they own. A workspace is you, chatting with your models. A team is agents working for you while you sleep — dispatched, reviewed, budget-capped, and interruptible. That's the next step. That's this.

Dogfooding

tester

The product is being built by its own agents.

The first deployment of agent.bytesmith is the team shipping agent.bytesmith. Every architectural decision has been stress-tested against real autonomous workloads. Nothing is speculative.

Pricing

Launch pricing. Waitlist members get 30% off any tier at launch.

Snapshot

^$99

early bird $69

One-time

"Read the code, own the architecture."

Versioned snapshot of the full system
As-is: no updates, no support
Complete source — no obfuscation
Self-deploy on your own VPS

Own your team before the price goes up.

This is pre-launch. No checkout today — just your place in line and 30% off any tier when we ship.

No SaaS. No vendor. No black box. Your VPS. Your model. Your agents. Your rules.

Your AI team. Your VPS. Your rules.