The AI control plane
for your fleet.
Describe a goal in plain language. Claude plans it, AI agents on each of your Linux, macOS, and Windows machines carry it out, and you approve anything sensitive. One goal, many machines, one result.
Concept preview. The fleet grid and tag filters are real; live per-machine progress during a goal is still in development. Machine names and output are illustrative.
How it works
From goal to done — automatically.
One real job, start to finish: rotate a TLS cert across five web nodes. Four things break along the way. The Director catches each one, repairs it, and finishes the job.
Illustrative run. You ask the Director to renew a TLS certificate that expires Friday on five web servers. The Director checks intent, loads policy and fleet context, writes a plan of audit, renew, reload and verify steps, classifies each by risk, checks scope and policy, and asks for your approval. Five operators then run the steps on their own machines. web-02 is missing certbot and asks permission to install it. web-03 fails an nginx config test, retries, and the Director rewrites the step. web-04 exits successfully but still serves the old certificate, which the Director's judge catches and repairs. web-05 runs out of disk and needs a typed confirmation to clean up. All five servers end up verified on the new certificate, and the Director returns one summary and writes the audit log.
Illustrative run. Hostnames, commands and output are examples, and approvals are condensed to one prompt per kind.
Features
Everything a fleet needs.
Any machine, any OS
Servers, VPSs, Proxmox and Docker hosts, laptops, and workstations, on Linux, macOS, or Windows. The Operator CLI runs anywhere Python runs.
Claude plans. Mistral executes.
Claude (Anthropic) handles planning and judgment. Mistral Small 4 handles fast, low-cost execution. All model calls happen on the Director, and your machines hold no AI keys.
Human approval gates
Sensitive operations — sudo, destructive commands, secret access — require your explicit approval before running.
Capability discovery
Each Operator probes its own tools at startup. The Director plans with full knowledge of what's actually available on each machine.
Rich context attachments
Attach screenshots, error logs, HTML, config files, or code. The Director sees everything before planning.
Append-only audit log
Every action is recorded with user, timestamp, and IP. Audit trail you can export anytime.
Security model
Your machines. Your data. Your rules.
Commands run on your hardware, not in cloud sandboxes. Task data travels to the Director and its model providers over TLS-encrypted connections. API keys never leave the Director. Multi-tenant isolation enforced at every query.
FAQ
Questions, answered.
Can AI agents coordinate work across multiple machines?
Yes, that's what DirectOps is for. Each machine runs a lightweight Operator agent. The Director (powered by Claude) splits your goal into tasks, picks which machines run what, and pulls the results together. For example, build on one host, test on another, or ask a question of your whole fleet at once.
How is this different from running Claude Code over SSH?
A single coding agent over SSH holds keys to every box and works one machine at a time. DirectOps runs an Operator on each machine that connects out to the Director, so there's no SSH key sprawl. Tasks run in parallel across machines, and every action passes through a central policy engine, approval gates, and an audit log.
Which AI models does DirectOps use?
Claude (Anthropic) for planning, task decomposition, and judging results. Mistral Small 4 for fast, cost-efficient command execution. All model calls happen on the Director, so your machines never hold AI API keys. Bring-your-own-key is supported.
Which operating systems are supported?
Linux, macOS, and Windows. That covers bare-metal servers, VPSs, Proxmox and Docker hosts, and workstations. Install the Operator with pip (directops-operator) or the one-line installer.
Is it safe to let an AI agent run commands on my servers?
A policy engine blocks dangerous commands. Changes show a dry-run preview first. Sudo, destructive, or secret-touching actions wait for your explicit approval, and everything is written to an append-only audit log. Fleet-wide questions are read-only by default.
Early access
Get early access.
We're onboarding teams in small batches — priority to DevOps and infrastructure teams running fleets of servers and workstations.