Open source · runs on your own computer
Let your agent manage your devices. You stay in charge.
Your agent already works in your files and your terminal. Hardpoint adds your devices: a checked card for each one says what it can read, what it may change, and what needs your yes first.
Bring your own agentClaude CodeCodexOpenCodeany LLM with a terminal
Ask about a device
Keep this change?
Made from the cards in the catalog. The answers are examples, not readings from your computer.
The difference
Why not let the agent work it out?
Claude Code, Codex or any LLM with a terminal can already try to control a device: it searches, writes a command and runs it. This is what changes with Hardpoint.
Knowing your device
Has never seen your exact device. It searches, then works from whatever it finds.
Reads a card written for that exact model.
Knowing the risk
Reading a battery and replacing firmware look the same to it: a command to run.
Every ability on a card has one of four labels, set before anyone asks.
Proof
Reports what it did. Nothing says whether that was ever checked on your model.
Each claim on a card says whether a person tested it on a real device, and links to the record.
Touching the device
Writes its own commands and runs them as you.
A card holds no commands. Hardpoint checks the label, then uses a known adapter or helper tool.
When it goes wrong
If a change locks you out, you are on your own.
A risky change is labelled to go back by itself unless you keep it. One that cannot be undone waits for your yes.
What it learns
Gone when the session ends. The next agent works it out again.
A card is checked once and shared. Everyone with that device gets it, in any agent.
Try a risky request
On its own
It does it straight away.
With Hardpoint
It happens. One person asked a voice assistant to "turn off the office", meaning the lights, and it also cut power to his server rack. His write-up.
Put it to work
More useful than another settings menu
Playbooks, watch configurations and workflow ideas—not preinstalled automations. Registry links show the related abilities; verify the route, permissions and operation before using one.
Workflow idea · integration required
Give your agents an attention budget
Let urgent decisions interrupt you while routine progress stays quiet. Keep a key associated with each session and queue lower-priority warnings. A session-focus or answer action needs an explicit host integration; a light alone is not an approval.
Related registry abilities: light.per_key
Configure a watch
Turn a spare key into a device inbox
Choose a reading and threshold that deserve your attention. Raise the warning on a configured spare key, then clear it when the reading recovers. Keep device warnings distinct from requests for permission.
Related registry abilities: battery.level · light.per_key
Agent playbook
Preflight the expensive job
Before a long local build, render or backup, ask your agent to collect the relevant readings and report blockers or unknowns. A drive's health verdict is not a failure forecast, and this checklist does not start the workload or prove a backup is safe.
Related registry abilities: battery.charging · thermal.pressure · storage.health
Workflow idea · integration required
Borrow a setup for a focused review
Propose a temporary precision, scrolling and display preset under each card's actual label. Build ownership and manual-override handling before automatic restoration, so ending a review never overwrites a setting you changed yourself.
Related registry abilities: pointer.dpi · scroll.mode · display.brightness
Workflow idea · integration required
Explain the missing connection
When a desk device disappears, trace which workflows depended on its connection instead of flooding you with unrelated warnings. This needs connection/topology integration; model or firmware readings alone cannot diagnose a cable or prove the dependency.
Related registry abilities: identity.firmware
Workflow idea · integration required
Make leaving a checklist, not a power cut
Review what is still in use, then propose a handoff checklist through your owning automation system. Treat every selected ability under its registry label. A broad request to shut down a room must never become blanket approval to interrupt power.
Related registry abilities: media.playing · power.outlet
Guided setup
Add the devices you choose
Discover what is present, choose all uncatalogued devices or specific entries, and have your agent reuse existing registry routes where they fit.
How it works
The labels
Every ability a device has gets one. The colour always means the same thing.
More than answers
Your devices can reach you, too
A card works both ways. In this illustration each agent session gets a spare key. Its light asks for your attention; a press runs the action you configure.
An interactive illustration drawn from a registry card and example configuration. You can click a lit key.
The catalog
Devices with a card
The bar shows how much of each card has been checked on a real device.
Questions
Fair things to ask
Does this let an agent control my computer?
Yes. That is what it is for: the computer itself and what is attached to it, such as the battery, the screen, the keyboard, the mouse and your drives. What changes is how. Your agent does not make up commands; it works from each device's card, where every ability has a label. Reading is always allowed. Small changes can be put back at once. Changes that could get in your way go back unless you keep them. Changes that cannot be undone, such as a firmware update, wait for your yes.
What can it do right now?
The device registry is the current source for abilities, labels and verification; the adapter registry supplies routes, platforms and helper requirements. These pages are generated from the registry rather than maintained as a second list here. On your computer, use hardpoint inventory, hardpoint doctor and the local MCP tools to check actual availability and errors.
Does it need any special permission?
Requirements belong to each card and adapter entry, not a separate permission list here. hardpoint doctor reports local helper requirements; actual operations report permission or connection errors. Hardpoint never grants permissions, elevates privileges, bootstraps a package manager or flashes firmware as helper setup. [tools] auto_install = false turns automatic helper setup off.
Does it replace my device or automation apps?
No. Hardpoint is a policy and verification layer over device tools, not a replacement for their specialist features. Consult the adapter registry to see which routes exist and their constraints. Keep one owner for each setting so two automation loops do not fight; research examples are not promises that an integration is available.
Which agents does it work with?
Use a compatible MCP client with the local stdio command hardpoint mcp. Configure it using your coding harness's current help rather than assuming a particular provider. Session events can be reported with hardpoint emit; a session-key press runs your configured action. Answering a host's question needs its own integration. Hardpoint device approvals remain a separate human decision.
What stops it doing something I did not want?
A card cannot supply code or an install command; helper installation uses trusted recipes built into the program. Every ability has a lowest label, so a card cannot make a risky change harmless. Gated actions wait for your approval of that exact action. These protections apply to requests through Hardpoint: an agent with unrestricted shell or device access can bypass it. Hardpoint is not a sandbox.
Is it watching what I type?
To let you answer with a spare key, it sees which key went down. It does not keep what you type. It keeps a count per key only if you turn that on.
Does anything leave my computer?
Hardpoint has no account or cloud service; the daemon listens locally. Catalog and program updates contact the source repository, and enabled helper setup contacts the package manager's registries. Readings you give your agent follow that agent's privacy settings and may be sent to its model provider. Shared cards contain model facts, not your unit's serial number, address or private name.
Does it cover smart-home devices?
Check the registry for matching models or class cards and the status of their adapter routes. A card's presence does not establish that control is available. Use the selected card's matching rules, requirements and verification, and keep private network identities out of shared data.
Will it work with my devices?
Start with discovery and the registry, then check the selected operation locally. The add-device prompt lets you choose all uncatalogued devices, all detected devices for review, or specific entries. Existing cards are reused or reviewed rather than duplicated. Your own cards can live in ~/.config/hardpoint/cards without being published.