nfltr.xyz runs Claude Code agents on the laptops, VMs and private servers that hold your code, data and test environments. Join each machine as a node and set what its agents may do; give the hub a goal, and it spawns agents where the data or access is, steers them, and brings back their answers and branches. The hosted relay carries messages and runs no model; Claude makes every decision.
Quickstart — join a machine, then give the hub a goal from your terminal or from your own Claude Code session:
nfltr node join --max-agents 2 --allow-all-tools
nfltr orch "fix the flaky cache test and push a branch with the fix"
claude mcp add nfltr -- nfltr mcp --toolset hubSee it run in the demos, or read the Agent Orchestration overview.
Install → API key → join machines → give the hub a goal → watch, steer, stop → recover.
The whole path in one page: install, API key, credentials, join machines, first goal, recover.
Guide 00nfltr orch "<goal>", the interactive hub, budget caps, hub ids and reattach, nfltr orch doctor, zero-setup local hub.
Machine routing, parallel work, watching tasks, and steering, continuing or stopping agents through the hub.
Guide 02API keys, shared nodes, one hub id per person or project, labels, and what teammates see.
Guide 04Symptom to fix: missing nodes, tools_not_allowed, node_lost, budgets, relay restarts, reattach, and recovery.
Make your own Claude Code session the hub: two commands, the completion hook, permissions.
Guidenfltr node join and nfltr orch nodes: tools, repositories, machine descriptions, credentials, isolation.
Step by step: nfltr relay serve on your Linux VM with your TLS certificate or a pinned self-signed one, one nfltr config set-relay step per machine, then nodes, goals and your own Claude Code; what works in the beta and what does not.
A small web service drives the hub over MCP, with no model in it: teams submit tasks over HTTP, webhooks or GitHub comments, Claude Code agents run on your machines under each one's policy, and results and costs come back. With a recorded developer's day.
TutorialWhat the hub is, how agents reach your machines, and what nfltr does and does not decide.
Corespawn_agent, send_message, stop_agent, list_agents, wait_for_agents, list_nodes: inputs, results, constraints, budgets, delivery guarantees.
Short answers: what nfltr orch is, how it compares, what the relay sees, whose Claude account runs the agents, and where to start.
GuideRecorded runs: two from a plain Claude Code session, one goal on two machines, a parallel backlog, a remote investigation, a node killed mid-job.
DemoStep-by-step tasks with the commands to type and the output to expect. Every step is tested against the CLI.
Give a reviewer a link to your localhost app, protect it with a generated password, revoke it.
Tutorialnfltr tail, nfltr watch and nfltr share for files, text and pipes.
nfltr shell or nfltr tcp 22, then ssh through nfltr ssh-config, and the host-key trade-off.
nfltr wg serve --allow on the LAN, wg socks5 or wg connect from anywhere.
One random key file on every machine keeps the relay out of the key exchange.
Tutorialp2p send, a2a send, notify send and the clipboard between your machines.
Which check first: diagnose, status, orch doctor, orch diagnose relay.