Cosmo

An assistant that knows your cluster, and asks first.

Cosmo lives in the panel. It reads your apps, their logs and their history to tell you what went wrong, and proposes the fix as a change you approve. It runs on the model you choose, with your key or on your own hardware.

Install Coswarm How it stays safe

Off until an administrator turns it on. The panel works fully without it.

How a change happens

It looks, then it asks.

1
Ask
In words, from any page, or with Explain where something broke.
2
Read
It uses its 31 read tools to look before it speaks. Reading needs no approval.
3
Propose
A change is a plan with no side effects: a diff, a risk level, what it will restart.
4
Approve
You say yes to that exact plan. Deletes need the name typed out.
5
Run
It runs with your permissions and is written to the change record.

panel.acme.dev/cosmo

checkout went down after the deploy. Why?

4 reads · 3 s

Read checkout: 0 of 2 replicas runningget_app
Read its tasks: 6 failed, “non-zero exit (1)”get_app_tasks
Read the last 100 log linesget_app_logs
Compared 2.2.4 with 2.3.0diff_app_versions

The new image reads its database address from DATABASE_URL, and checkout never had it: 2.2.4 built it from DB_HOST, DB_USER and DB_PASSWORD. Every replica exits at start with DATABASE_URL is not set.

I can put the same address together from the values it already has:

Set an environment variable on checkout

set_app_env

Medium risk

env DATABASE_URL

postgres://checkout:•••@postgres:5432/shop

checkout restarts with a rolling update; 2 replicas, one at a time.

Expires in 4:52RejectApprove

Approve or reject the proposed change, or write something else.

Where you meet it

Not a chat box in a corner.

Cosmo is wired into the places where questions come up.

On the dashboard

Each finding the server spots opens its own conversation. Cosmo’s answer becomes the “why”, its proposal the “fix”, approved on the spot.

Where something went wrong

A rollout that stalled, a container killed for memory, an error line, an incident: each has an Explain that opens Cosmo with the context filled in.

Wherever you are

A side panel on every page with ⌘J, its own page for long work, and “Ask Cosmo” in the command menu when a search finds nothing.

While you are away

Watches check on something every 15 minutes to weekly, or when an app goes down, and tell you in the panel or in Slack, Discord or Mattermost.

What it can do

51 tools, and a short list of things it won’t.

Reads run on their own. Every write shows an approval card with its risk; low-risk ones can go ahead without asking, per conversation, if an administrator allows it.

Reads 31 · no approval

Apps

list_appsget_appget_app_tasksget_app_logsget_app_statssample_app_statslist_incidentsapp_timelinediff_app_versions

Cluster

get_swarm_overviewlist_containersget_containerget_container_logsget_disk_usage

Images

list_local_imagessearch_imageslist_image_tagsread_image_docscheck_image_updates

Platform

list_projectslist_volumeslist_networkslist_registrieslist_secretslist_domainsget_trafficlist_backupslist_backup_fileslist_app_configslist_approvalsopen_in_panel

Changes 20 · each one approved

Low risk

restart_appprobe_from_apptake_backup

Medium risk

scale_appupdate_app_resourcesupdate_app_portsattach_app_networkattach_app_volumeset_app_envrollback_appcreate_appinstall_composeadd_app_domainupdate_app_domainremove_app_domainprune_imagesdelete_backup_file

High risk

deploy_imagedelete_appdelete_volume

On purpose, never

No shellNo SQLNo writing filesNo users or rolesNo arbitrary HTTP

Safety

Built so that a wrong answer can’t become a wrong change.

A model can be mistaken, or misled by what it reads. Cosmo is built around that.

Secrets stay on the server

Private keys, tokens, cloud and API keys, JWTs and password variables are masked before anything reaches the model. A value you paste becomes a placeholder the server fills in when the change runs.

What it reads is data, not orders

Logs and tool output are marked as untrusted, so a line in a log that says “ignore your instructions” stays a line in a log.

Your permissions, twice

Cosmo can do only what your role can. Permissions are checked when the plan is made and again just before it runs. It can never approve anything itself.

A yes that expires

An approval is a single-use token for you alone, bound to a hash of the plan, valid for ten minutes, five for high risk. If the app changed in between, nothing runs.

Four eyes stay human

In areas held for a second person, Cosmo does not make the change. It takes you to the page in the panel where the request goes through review.

Limits you set

Messages per minute per user, daily token budgets per user and in total, a read-only mode, and how long conversations and the change record are kept: 30 and 180 days by default.

Bring your own model

Your key, or your own hardware.

Add a provider in Settings; it is checked to make sure the model can call tools before anyone uses it. Keys are stored encrypted and never come back to the browser. A helper model can write titles and summaries, and a second provider can step in when the first is down.

OpenAI

Responses API, nothing stored

Anthropic

Extended and adaptive thinking

OpenRouter

Any model it carries

Ollama

On your hardware, no key

OpenAI-compatible

vLLM, LM Studio, LiteLLM…

The conversation

A chat that keeps up with the work.

Edit a message or write a reply again, and move between versions
Slash commands and @ mentions of apps
Files and pictures attached to a message
Charts and live app cards in the answer
A plan it keeps for work in several steps
Subtasks in conversations of their own
Memory per user, and the administrator’s instructions
Replies run on the server and survive a reload
Pin, export as Markdown, share to read
Model and effort chosen per conversation

Let Cosmo meet your servers.

Install Coswarm, add a model in Settings, and ask it what is going on.

curl -fsSL https://coswarm.dev/install.sh | bash

See the panel Meet Cosmo