Boot and run AI agents that live on their own Linux server, with their own IP, email and memory.
Scored across 21 tools
Each tool targets a fairly distinct resource+action: machines (add/list/remove/test), agents (boot/get/set_model/set_temper), duties (get/save/run), files (read/write), secrets, logs, manual. The main soft spots are talk_to_agent vs get_reply (send-and-maybe-poll vs fetch-result) and get_agent vs get_account vs get_duties, which overlap in what they report but are distinguished by scope in their descriptions.
Every name follows a strict snake_case verb_noun pattern: get_*, read_*, set_*, save_*, add_/remove_*, run_*, boot_*, talk_to_*. No mixed conventions or vague standalone verbs.
21 tools is on the heavy side of the ideal band, but the domain genuinely spans machines, agent lifecycle, duties, files, secrets, account, and docs, so nearly every tool earns its place. Minor compression possible (e.g. read_logs/read_manual as read modes), but nothing looks gratuitous.
Machines and duties and files have near-full lifecycles, but there is no way to delete a duty (only pause via run_duty) and no agent teardown/cancel despite boot_agent incurring a recurring $9/month charge. Secrets can be written but never listed or removed, leaving a few real dead ends.
Adds or updates a server an agent may log into with its own SSH key (the server needs a login for that key first; read_manual which=api has the recipe). watch: true includes it in an immune-system agent's sweeps.
Inputs: host (required), user (required), alias (required), agent_id (required), +3 more.
Boots a new agent on its own server. It costs $9 a month from the bootkey balance, billed by the hour, plus model tokens on the account's own key, and is ready in about two minutes (get_agent shows its progress).
Inputs: name (required), model (required), region (required), provider (required), +5 more.
Returns a one-click link where the person connects their own OpenRouter account to bootkey (the key goes straight to bootkey and never passes through the chat), or where to paste a Venice key. An account needs a model key before an agent can boot.
Inputs: provider.
The bootkey account: balance and what booting another agent needs, connected model keys, SSH keys, every agent with its status, and what this connection may do.
Inputs: no named parameters.
One agent: status, model, temper, web address, email, what it is doing now and its next wake-up, its latest line, vault names and recent activity.
Inputs: agent_id (required).
An agent's duties (scheduled jobs with schedule, next run and who set them), its model, Hermes version and skills.
Inputs: agent_id (required).
The jobs (templates) an agent can start with, and model presets with prices per million tokens and leaderboard ranks.
Inputs: no named parameters.
The result of an earlier message or command to an agent, waiting up to 50 seconds for it.
Inputs: agent_id (required), command_id (required).
The servers an agent may log into over SSH (alias, host, user, port, note, watched, fingerprint) and the public key it logs in with.
Inputs: agent_id (required).
Lists the files in an agent's ~/.hermes (no path), reads one file (path), shows its saved versions (path and history: true), or one old version (path and version).
Inputs: agent_id (required), path, history, version.
The last lines of an agent's system journal: hermes (the agent) or agentd (bootkey's daemon on its box).
Inputs: agent_id (required), unit, lines.
The bootkey manual (every option with its trade-offs: temper, memory, regions, costs, access, vault, email, Nostr) or the API reference.
Inputs: which.
Removes one of the servers an agent may log into.
Inputs: alias (required), agent_id (required).
Runs one of an agent's duties now, pauses it, or resumes it.
Inputs: id (required), action (required), agent_id (required).
Adds a duty (a scheduled wake-up with its own instructions) to an agent, or changes one when id is given. schedule: five-field cron ("0 8 * * 1-5"), an interval ("30m", "every 2h") or a one-off ISO time ("2026-12-01T09:00").
Inputs: prompt (required), agent_id (required), schedule (required), id, +2 more.
Saves API keys or tokens to an agent's vault as NAME: value pairs. They are encrypted, can't be read back, and reach the agent within seconds.
Inputs: vars (required), agent_id (required).
Switches an agent's main model to another one from the same provider (ids from get_options).
Inputs: model (required), agent_id (required).
Changes an agent's temper: reactive (tool), pulse (worker, wakes every N minutes all day), daynight (worker by day, dreams at night) or self (alive, picks its own wake-ups). Its other duties stay.
Inputs: kind (required), agent_id (required), every.
Sends a message to a running agent, as its human would in the console chat, and returns its reply. Long tasks can take minutes; then it returns a command id for get_reply.
Inputs: message (required), agent_id (required).
Has an agent log into one of its servers and returns the result and the server's fingerprint. On the first login the agent remembers that fingerprint.
Inputs: alias (required), agent_id (required).
Writes a file that shapes an agent's job: SOUL.md, objectives.md, now.md, or a file under watch/, notes/ or immune/ in its ~/.hermes. The previous version stays in its history.
Inputs: path (required), content (required), agent_id (required).
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Persistent Linux computers for AI agents: desktop, signed-in browser, vault, apps and shell.
Real Linux labs your AI agent deploys, routes and runs, with domains, TLS, DBs and an audit log.
Cloud desktops for AI agents: create a Linux computer, see its screen, click, type and run tasks.
Your own cloud computer run by an AI agent: signed-in browser, its own email, files, long jobs.
Add one secure layer between your agents and this server.
