Host the agents you already run, right next to your app.
Lucity runs the agent from your terminal as it is, around the clock. It wakes on the events you choose, reaches only what you allow and keeps a record of every call. A cloud platform made in Switzerland and hosted in the EU.
Awake for 11 minutes. Asleep, and nearly free, the rest of the day.
Deploysapi, release 214
Lucity
The deploy of api failed while building.
Deploy doctor
On it. Reading the build log and the last three commits.
Deploy doctor
The build needs STRIPE_WEBHOOK_SECRET since commit 4f2a91c, and it is not set at build time. The previous release still serves traffic. Should I add the variable from the shared group and redeploy?
Add and redeployNot now
Asleep when idleNo work, no bill worth mentioning. A trigger wakes it in seconds.
Picks up where it left offFiles, memory and conversations survive restarts, deploys and sleep.
A sandbox of its ownEach agent runs isolated, on infrastructure built for other people's code.
Answers where you talkTelegram, Mattermost, Slack, Teams or the dashboard, approvals included.
Under your control, down to a single tool.
Every agent starts with nothing. You hand it each model, tool and host on purpose, the keys never go inside it, and you can read back everything it did.
Night watch · access
Model: Mistral Largeallowed
Lucity: logs, metrics, deploysallowed
Lucity: memory, instances, rollbacksasks Ops first
Lucity: delete anythinghidden
Sentry: read issuesallowed
Linear: create issuesasks first
Internet: status.mistral.aiallowed
Network: other services in productionoff
Model spend per day€20, then it stops
RecordNight watch · last night
03:12:41triggerrestart loop · webwoke
03:12:44modelmistral-large · 4.1k in, 380 outok
03:12:52toollucity · logs, metricsok
03:13:11toolsentry · read issuesok
03:13:20toollucity · set memory to 1 GBheld
03:15:02approvalNoa, on Telegramallowed
03:15:03toollucity · set memory to 1 GBok
03:15:44networkpaste.example.orgblocked
Shop on call · 03:13
Night watch wants to set the memory of web to 1 GB.
Changes to production need a yes from someone in Ops. The agent cannot answer for itself.
ApproveDeny
Grants, one by oneAllowed, asks first or hidden, down to a single tool of a single connector.
Keys stay outsideLucity adds the key as a call leaves. The agent never holds one, so it cannot leak one.
A yes from the right personRisky calls wait. You decide who may answer, and they answer from wherever they are.
A record of every callModel calls, tool calls and approvals, kept for a year, searchable and exportable.
Also included: a list of hosts each agent may reach, a daily spend limit, step by step replay of any run, and export of the record to your own bucket. How agent security works
Built from parts, not from a framework.
An agent is a harness you know, instructions in plain words and four kinds of parts. Connect a part once, and every agent in the workspace can use it when you allow it.
Pick the agent you already run in your terminal and paste its instructions file. That is the whole move.
Open harnesses, pinned and kept up to date for you
One plain text file of instructions, the one you already have
Skills and subagents come along
Export an agent and run it anywhere a container runs
An event, a schedule or a message starts a conversation. Events about the same thing land in the same one, so a flapping service is one thread, not forty.
From Lucity: failed deploys, crash loops, full disks, new vulnerabilities
From your tools: alerts, new issues, new errors
Any schedule, in your time zone
A signed webhook, or an email to the agent's own address
An agent lives in an environment, next to the services it looks after. Try a change in staging, promote it to production, roll it back when it misbehaves.
Right next to your appAn agent reaches the services of its environment by name, when you let it. No tunnels, no public endpoints.
Every change is a versionInstructions, grants and triggers are versioned together, with who changed what.
One click to productionPromote the agent you tried in staging, exactly as it was.
shop/production
Night watchwakes on alerts for webOpenCode · v13
Asleep
webshop.examplezeitlos/shop
Online
ordersPostgreSQL2 instances
Online
Night watch · versions
v14stagingAlso watch the worker queueNoa · 2 hours agoPromote
v13productionAsk Ops before any rollbackChristian · Tuesdaylive
v12earlierReport to Shop on callNoa · 2 NovemberRoll back
What it costs, and where it runs.
Billed while awakeAn agent is billed like a service, by the memory and CPU it uses while it works. Asleep, it costs next to nothing.
Models at costModel calls go on your own provider bill, at their price. We add no markup and resell no tokens.
In EuropeAgents run in the EU, operated from Switzerland. Pair one with a European model provider and its work stays in Europe.
Nothing of ours insideThe harness is open and the instructions are a text file. Export an agent and run it wherever a container runs.
No. Pick the harness you use locally, paste its instructions and connect the same tools. The agent behaves the way it did in your terminal, minus the terminal.
Claude Code, Codex, OpenCode and Pi today, and the list keeps growing. If it runs in a terminal and reads an instructions file, it is a candidate. Tell us which one you are missing.
No. A held call waits outside the sandbox, and only the people your rule names can answer it, from the chat or the dashboard. The agent sees the answer, never the button.
The agent sleeps and costs next to nothing. An event, a schedule or a message wakes it in seconds, with its files and conversations where it left them.
Yes. It drives Lucity through the plugin as before. Hosted agents take the work that should not wait until you are back at your desk.
Any provider your workspace connects, from Anthropic and OpenAI to Mistral, Infomaniak or a model you host yourself. Each conversation can pick its own model and effort.
In the EU, inside your workspace, operated by a Swiss company. Pair it with a European model provider and nothing the agent does leaves Europe.
Move one agent off your laptop.
1
Pick its harness
The one you run in your terminal today.
2
Paste its instructions
The instructions file you already have works as it is.
3
Tell it when to wake
A deploy fails, an alert fires, or every weekday at 06:00.
A trigger starts a conversation with an agent. It can be something that happens on Lucity or in one of your tools, a schedule, a webhook or an email. The agent sleeps until one fires and is at work seconds later.
Agent hosting
Triggers
A trigger starts a conversation with an agent. It can be something that happens on Lucity or in one of your tools, a schedule, a webhook or an email. The agent sleeps until one fires and is at work seconds later.
web · production40 events
Restart loop
06:02Restarted, out of memory at 512 MiB
06:04Restarted, out of memory
06:09Restarted, out of memory
and 37 more, in the same thread
Night watch is on it
Forty restarts, one thread
One thread per incident
Events about the same thing land in the same conversation. A service that restarts forty times in ten minutes opens one thread, and every restart after the first is added to it instead of waking the agent again. Lucity events group by service and kind, webhooks and emails set their own key.
What the agent is told
The event arrives as the first message, in plain words with the details attached: the service and environment, the last lines of its logs, the deploy and its commit, and a link back to the dashboard.
Quiet hours and limits
Each trigger can be paused, kept quiet outside the hours you set and limited to a number of runs per hour. A trigger that fires while its limit is spent is added to the thread and waits there.
Every trigger
SourceEvents
Lucity
Deploy failed, restart loop, out of memory, disk almost full, new vulnerability, certificate failed, backup failed