kasl-server
Team server for kasl. Employees run kasl on their machines; the agents send work-time data to the server. Managers get dashboards, charts, and reports across the whole team; every employee gets a personal page.
Status: pre-alpha. The door for kasl agents is open and survives a bad connection: a day at a time on
POST /api/v1/days, a backlog on/days/batch, and a task the employee deleted can be deleted here too. History from before the server arrived can be imported from an agent's own database; people sign in, and an administrator manages the team, its departments and its agent tokens without touching the host, and every such change is recorded. What the server keeps about a person is now a policy it enforces rather than a claim: an employee can ask it, and an administrator can narrow it. The loop is closed: agents deliver, employees see their own week, and a manager sees their team's - hours per person, a drill-down into anyone's days, and a live column saying who is working, who is on a break and whose machine has gone quiet, the month as a grid where a square is a person's day, and a band that names who is worth a look and why. The same binary serves all of it, and installing it is a compose file and a published image rather than a build. What is still missing is the other half of the loop: kasl cannot send on its own yet, so history is imported or posted by hand.
Try it
Requires Rust and Docker.
$ git clone https://github.com/lacodda/kasl-server && cd kasl-server
$ docker compose up -d db
$ export DATABASE_URL=postgres://kasl:kasl@localhost:5433/kasl
$ export KASL_AGENTS=employee@example.com:agent-token
$ cargo run
2026-08-29T18:15:18.767453Z INFO kasl_server: database schema is up to date version=20260830000001
2026-08-29T18:15:18.811231Z INFO kasl_server::provision: provisioned agents from KASL_AGENTS agents=1
An administrator account was created, because this installation had none:
email: admin@kasl.local
password: ye4e9vgapwi8ptrh9kt9
This is the only time it is shown. Sign in and change it.
2026-09-01T21:59:49.005920Z INFO kasl_server: kasl-server listening version="0.19.0" addr=0.0.0.0:8080 max_batch_days=31 max_body_bytes=4194304
$ curl http://127.0.0.1:8080/health
{"database":"ok","demo":false,"status":"ok","version":"0.19.0"}
# The web UI is served by the same binary on the same port - open
# http://127.0.0.1:8080 and sign in.
# An agent back from three days offline. The middle day is impossible - it ends
# before it starts - and the others land anyway.
$ curl -X POST http://127.0.0.1:8080/api/v1/days/batch \
-H "Authorization: Bearer agent-token" -H "Content-Type: application/json" \
-d '{"days":[
{"date":"2026-08-15","started_at":"2026-08-15T09:04:00-03:00","ended_at":"2026-08-15T18:12:00-03:00",
"tasks":[{"agent_task_id":7,"recorded_at":"2026-08-15T18:10:00-03:00","name":"Reliable ingest","completeness":60}]},
{"date":"2026-08-16","started_at":"2026-08-16T19:00:00-03:00","ended_at":"2026-08-16T09:00:00-03:00"},
{"date":"2026-08-17","started_at":"2026-08-17T09:11:00-03:00","ended_at":"2026-08-17T17:40:00-03:00",
"tasks":[{"agent_task_id":7,"recorded_at":"2026-08-17T17:38:00-03:00","name":"Reliable ingest","completeness":100}],
"tasks_are_complete":true}]}'
{"accepted":2,"rejected":1,"results":[
{"status":"accepted","workday_id":"82ca500d-feeb-4d1f-8fb7-0b376339be02","date":"2026-08-15","pauses":0,"tasks":1,"deleted_tasks":0},
{"status":"rejected","date":"2026-08-16","error":"ended_at is before started_at"},
{"status":"accepted","workday_id":"2ea46d40-3aa0-48d3-8d8d-e1bb152a36bc","date":"2026-08-17","pauses":0,"tasks":1,"deleted_tasks":0}]}
# The employee deletes the task in kasl; the agent re-sends the day and says so.
$ curl -X POST http://127.0.0.1:8080/api/v1/days \
-H "Authorization: Bearer agent-token" -H "Content-Type: application/json" \
-d '{"date":"2026-08-17","started_at":"2026-08-17T09:11:00-03:00","ended_at":"2026-08-17T17:40:00-03:00",
"tasks":[],"tasks_are_complete":true}'
{"workday_id":"2ea46d40-3aa0-48d3-8d8d-e1bb152a36bc","date":"2026-08-17","pauses":0,"tasks":0,"deleted_tasks":1}
The task is gone from the 17th - and still there on the 15th, where the employee did not delete it.
The dev database listens on 5433, leaving a PostgreSQL you may already run on
5432 alone; override with KASL_DB_PORT.
The demo
A fresh server has nothing on it, and an empty dashboard is where most trials
end. KASL_DEMO=true fills an empty database with a fictional team — three
departments, twelve people, eight weeks of days — so the dashboards can be seen
before an agent is installed on anybody's machine:
$ KASL_DEMO=true docker compose up -d # or: KASL_DEMO=true cargo run
$ docker compose logs server
2026-08-29T18:14:37.667958Z INFO kasl_server: database schema is up to date version=20260830000001
2026-08-29T18:14:39.401685Z INFO kasl_server: seeded the demo team people=12 departments=3 days=386
This is a demo: a fictional team, nothing here is real. Sign in as
manager priya.raman@example.com Priya Raman
employee tomas.verhoeven@example.com Tomas Verhoeven
admin sam.whitfield@example.com Sam Whitfield
with the password `kasl-demo`. The same password opens every account.
$ curl http://127.0.0.1:8080/health
{"database":"ok","demo":true,"status":"ok","version":"0.19.0"}
The login screen offers the same three accounts as buttons, and every screen carries a banner saying the data is invented.
The days are shaped so that everything the dashboard knows how to show is on it at once: someone steady, someone working ten-hour days, someone whose hours shrink week by week, a day open right now, an agent that went silent a week ago, one that never reported, and an administrator with no agent at all. The history ends yesterday whenever you start it, and two demos started on the same day show the same numbers, so a screenshot can be reproduced.
The live status is on it too: somebody working, somebody on a break, agents idle between days, and two machines that have stopped answering. A demo is seeded once but a pulse is believed for three minutes, so the demo re-stamps its own — keeping the stopped agents stopped, and leaving alone any real kasl pointed at it. A demo seeded before the pulse existed gets its own on the next start, so upgrading the image is enough.
The demo refuses a database that already holds accounts. Twelve invented people alongside a real team, with nothing to say which rows are which, is the one outcome worse than no demo — so a flag left in a file after a trial stops the server with a message rather than turning the installation into one. A database the demo itself seeded starts normally, with or without the flag, and keeps its banner: the mark lives in the database, not in the environment.
The agents' tokens are demo-<firstname> — demo-tomas, for instance — so a
real kasl can be pointed at the demo and its days appear next to the invented
ones. Every address is under example.com, which is reserved: the names are
made up, and the domain guarantees the addresses are too.
The API
/api/v1 from the first endpoint: agents update on their own schedule, so a
path keeps meaning what it meant when the agent calling it shipped.
POST /api/v1/days — upload one day. Requires Authorization: Bearer <token>. The body is the workday with its pauses and tasks; ended_at is
absent while the day is still running, and so is a pause's, and tasks may be
empty.
Two properties are worth knowing before writing a client:
- Every instant needs a UTC offset, and the day carries its own
date.2026-08-14T09:12:00without an offset is refused (422): one team's hours have to stay comparable across time zones, and which calendar day work belongs to is the agent's call, not a value derived on the server. - The last upload wins. Re-sending a day replaces what is stored, so a
correction made in kasl lands and a retry after a lost connection is safe -
the same payload twice leaves the same rows. Pauses are replaced as a set;
tasks are matched on
agent_task_id, so a task carried into the next day moves rather than multiplying. - Deleting a task takes one word. Send
"tasks_are_complete": trueand the date's tasks the payload omits are deleted, which is how a task the employee removed in kasl disappears here too. Other dates are untouched. Leave the flag out - as agents written before it did - and nothing is ever deleted.
POST /api/v1/days/batch — upload a backlog. The body is {"days": [...]}
with the same day objects, and the answer reports each one:
Each day is written on its own, so one the server will never accept does not
block the rest - an agent that could not deliver any of its backlog because of
a single bad row would retry the same request forever. The batch carries at most
KASL_MAX_BATCH_DAYS days (31) and the body at most KASL_MAX_BODY_BYTES
(4 MiB); past either, 413.
Which failures are worth retrying. 4xx means the payload will not be
accepted as sent, however many times it is tried - fix it or drop it. 5xx
means the fault is on this side; send it again later. A batch that answers 5xx
stopped partway: the days already accepted are stored, and re-sending them is
safe because the last upload wins.
GET /api/v1/agent/whoami — whose token this is. An agent token is opaque
to the machine holding it, so a token pasted from the wrong place would file
that machine's days under a colleague's name without a word. This answers the
question at connect time, while a person is watching:
$ curl -H "Authorization: Bearer $KASL_TOKEN" http://127.0.0.1:8080/api/v1/agent/whoami
{"user_name":"kirill","agent_name":"laptop","api_version":"v1","server_version":"0.14.1"}
Refused with 401 for a token that is unknown, revoked, or belongs to a
deactivated account - the same answer the upload routes give, because a token
that cannot write must not be able to read a name off the installation either.
Every accepted day reports the privacy level that applied, and a
discarded object when that level left something out. See
what the server stores about you: an agent
should never report a break as recorded by a server that did not store it.
A malformed day is refused with 400 and a reason naming the field
({"error":"tasks[0]: completeness must be between 0 and 100"}); an
unrecognized, revoked or deactivated token gets 401. The reasoning behind all
of this is in ADR 0004
and ADR 0005.
Signing in
People sign in with an email and a password; kasl agents keep using their bearer token and are unaffected by any of this.
$ kasl-server admin --email boss@example.com --password '...'
admin boss@example.com is ready
$ curl -i -X POST http://127.0.0.1:8080/api/v1/auth/login -H "Content-Type: application/json" -d '{"email":"boss@example.com","password":"..."}'
HTTP/1.1 200 OK
set-cookie: kasl_session=b7857303342e1a4b...; Path=/; HttpOnly; SameSite=Strict; Max-Age=1209600
{"status":"ok"}
$ curl -H "Cookie: kasl_session=b7857303342e1a4b..." http://127.0.0.1:8080/api/v1/auth/me
{"id":"de836432-5dce-4705-9344-a65b356fc662","email":"boss@example.com","display_name":"boss@example.com","role":"admin"}
POST /auth/logout ends this session, POST /auth/logout-everywhere ends all of
them, and GET /auth/me says who the caller is.
Sessions are rows in the database, not signed tokens. The query per request buys the thing a self-contained token cannot give: access ends when it is ended — the afternoon someone leaves, not whenever their token happens to expire. A session lasts a fortnight and each use pushes that out again.
An unknown email, a deactivated account and a wrong password all answer
{"error":"wrong email or password"}, so the login form cannot be used to find
out who works somewhere.
The first administrator comes from kasl-server admin or
KASL_ADMIN=email:password in the environment. Running it again resets the
password and promotes the account, which is both the way back in after a
forgotten password and the way to make an admin of someone whose account already
exists because their agent has been reporting. Accounts created by KASL_AGENTS
have no password and cannot be signed into — they exist to own an agent's data.
Set KASL_SECURE_COOKIES=false when serving over plain http://. A Secure
cookie is silently dropped by the browser there, which looks exactly like login
doing nothing. The reasoning is in
ADR 0007.
Reading your own days
GET /api/v1/me/days answers the signed-in person's own history: the workdays
in a range, each with its pauses and the tasks logged on it. Both ends are
inclusive, so one date twice is one day.
$ curl -H "Cookie: kasl_session=..." "http://127.0.0.1:8080/api/v1/me/days?from=2026-08-24&to=2026-08-30"
{"from":"2026-08-24","to":"2026-08-30","privacy_level":"full","not_stored":[],
"days":[{"date":"2026-08-24","started_at":"2026-08-24T12:05:00Z","ended_at":"2026-08-24T21:12:00Z",
"worked_seconds":29640,"paused_count":2,"paused_seconds":3180,
"pauses":[{"id":"56e75449-...","started_at":"2026-08-24T15:30:00Z","ended_at":"2026-08-24T16:15:00Z",
"duration_seconds":2700,"manual":true,"reason":"lunch"}],
"tasks":[{"id":"2a9bd688-...","name":"Read-only endpoint","comment":null,
"completeness":100,"recorded_at":"2026-08-24T20:50:00Z"}]}]}
worked_seconds is the day's span minus what was paused, and it is null while
the day is still open — a day in progress has no total, and reporting the hours
so far as the day's figure would make every working afternoon look short.
paused_count and paused_seconds are always answered, whether they come from
the stored pauses or from the totals a coarse policy keeps instead.
The route is /me, not your own id under /users. It consults no role and
no department, so there is no permission here to read wrong; reading someone
else's days arrives with the manager's dashboard as its own route, where the
check is the point. A session is required — an agent's bearer token writes days
and reads the privacy manifest, and that is deliberately the whole list.
not_stored names what the installation's privacy level withholds —
pauses, tasks, free_text, or nothing at all. It is what lets a screen say
"not stored" where it would otherwise draw an empty section: an employee cannot
tell "no pauses were kept" from "you took no breaks", and only one of those is
true (ADR 0011).
A range covers at most 400 days; a wider one, or one that runs backwards, is a
400 naming the span asked for rather than a truncated answer.
The team's hours
GET /api/v1/team/days answers a row per person over a range: hours, days
recorded, pauses, and what the server knows about them right now. Managers and
administrators only.
$ curl -H "Cookie: kasl_session=..." "http://127.0.0.1:8080/api/v1/team/days?from=2026-08-24&to=2026-08-30"
{"from":"2026-08-24","to":"2026-08-30","privacy_level":"full","not_stored":[],
"members":[
{"id":"c49ea6a8-...","display_name":"Anna","email":"anna@example.com","department":"Engineering",
"days_recorded":3,"worked_seconds":72600,"paused_seconds":8100,"last_day":"2026-08-26",
"day_open":false,"last_seen_at":"2026-08-26T20:31:00Z","agents":1},
{"id":"0073460d-...","display_name":"Clara","email":"clara@example.com","department":null,
"days_recorded":0,"worked_seconds":0,"paused_seconds":0,"last_day":null,
"day_open":false,"last_seen_at":null,"agents":0}]}
Everyone the reader may see is listed, including people with nothing recorded. Clara above has no agent installed and no days; she is on the list anyway, because an employee whose agent never reported is exactly who a manager needs to notice. A table that dropped her would hide the case it exists for.
day_open and last_seen_at are about days, not about this moment -
whether a day is open on that person's own calendar, and when one of their
agents last delivered anything. Who is at their keyboard right now is a
separate question with a separate endpoint, below.
Who sees whom is the rule departments established: an administrator sees everyone, a manager sees the departments they run plus themselves, and a person in no department is visible to the administrator alone (ADR 0009).
GET /api/v1/users/{id}/days is the drill-down: the same response as
/me/days, for a person the caller is entitled to see. An id they may not
see answers 404 rather than 403 - a manager probing ids should not be able
to tell an employee in another department from one who does not exist.
The month as a shape
GET /api/v1/team/heatmap?month=YYYY-MM answers the same team a day at a time:
a row per person, a cell per date they recorded. It is kasl sum widened by
one axis - what the week's totals cannot show is the pattern, and the pattern
is how a manager notices somebody working weekends or filing nothing since the
12th.
$ curl -H "Cookie: kasl_session=..." "http://127.0.0.1:8080/api/v1/team/heatmap?month=2026-08"
{"month":"2026-08","from":"2026-08-01","to":"2026-08-31","busiest_seconds":34200,
"rows":[
{"user_id":"c49ea6a8-...","display_name":"Anna","department":"Engineering",
"worked_seconds":72600,"busiest_seconds":34200,
"days":[{"date":"2026-08-24","worked_seconds":27000,"open":false},
{"date":"2026-08-25","worked_seconds":34200,"open":false},
{"date":"2026-08-26","worked_seconds":null,"open":true}]},
{"user_id":"0073460d-...","display_name":"Clara","department":null,
"worked_seconds":0,"busiest_seconds":null,"days":[]}]}
A date that is not in the row was not recorded. The server does not
manufacture a zero for it. Clara's empty days and a day of no hours are
different facts, and a grid that painted them the same shade would say an
employee who never installed kasl took the month off - the reading a manager
reaches for first, and the false one (ADR 0011). The web UI draws the absence
as an empty square and says so in the legend.
An open day carries worked_seconds: null, exactly as /me/days answers
it. A day half-lived is not a short day.
The scale is the data's own. The response carries seconds and each row's busiest day, plus the busiest anywhere in the grid; how those become colour is the screen's business. A threshold shipped in the API would be this server asserting what a normal working day is, and it has no opinion on that until norms arrive with the production calendar (ADR 0015).
month must be YYYY-MM. A date, a bare year, or a missing parameter is a
400 rather than a guess - a caller who sent one would otherwise get a
plausible answer to a question they did not ask. Visibility is the same rule as
every other team endpoint, and a person with nothing recorded is listed anyway.
Where to look
GET /api/v1/team/signals answers the question a manager does not know to
ask. A three-week slide exists only across weeks; no single screen shows it,
and nobody scrolls back through weeks hunting for one.
$ curl -H "Cookie: kasl_session=..." http://127.0.0.1:8080/api/v1/team/signals
{"from":"2026-06-08","to":"2026-08-30","people":12,
"signals":[
{"user_id":"0073460d-...","display_name":"Jonas Petit","department":"Design",
"kind":"no_data","days_quiet":13,
"weeks":null,"from_seconds":null,"to_seconds":null,"median_seconds":null},
{"user_id":"c49ea6a8-...","display_name":"Lukas Brandt","department":"Engineering",
"kind":"declining","weeks":3,"from_seconds":97200,"to_seconds":79560,
"median_seconds":null,"days_quiet":null}]}
Everything here compares a person with themselves. Never with a colleague and never with a norm - this server has none until the production calendar (v0.21), and a threshold invented before then would be this product asserting what a working day should be on somebody else's team. A dashboard that ranked people against each other would be a scoreboard, which this deliberately is not.
A signal is a question, not a verdict. Each one carries the figures it was computed from, so a screen says "27 h a week down to 22 h" rather than showing a badge that reads "problem". Falling hours are a holiday, a hospital, or a project that ended, and the server knows none of that.
The three:
declining— the last three weeks sit at least 15 % below the three before them, comparing the median of each side. Levels, not a run of falls: a genuinely fading person goes 33 → 24.8 → 27 → 20.9 → 23.1 → 22.1, which is an unmistakable slide and never three consecutive drops. One ordinary week resets a run, so counting steps stays silent on exactly the case this is for (ADR 0016).no_data— nothing recorded for ten days or more. The live status (below) only sees a silent machine once a pulse is expected; an agent too old to send one is invisible there.unusual_week— the last complete week is more than 40 % away from that person's own median, in either direction. Both ways on purpose: flagging only the short weeks would make the signal an accusation rather than a question.
The current week never enters the arithmetic. A Tuesday is not a short week, but that is what a partial week looks like to a sum - and including it would flag the whole team every Monday morning. Medians rather than means throughout, so one crunch week cannot hide a decline behind it.
GET /api/v1/users/{id}/trend is what a signal links to: twelve complete
weeks, the empty ones included, plus that person's median and the signals
about them. A week nobody worked keeps its place — closing the gap up would
turn an absence into continuity.
Nothing is stored and nothing is sent. The signals are a function of the workdays already in the database, and delivery — alerts, webhooks — is its own milestone.
Who is working now
An agent reports in every minute with what it sees:
$ curl -X POST -H "Authorization: Bearer $KASL_TOKEN" -H "Content-Type: application/json" \
-d '{"state":"working","at":"2026-08-30T14:22:10-03:00"}' \
http://127.0.0.1:8080/api/v1/agent/heartbeat
{"interval_seconds":60,"stale_after_seconds":180,"state":"working","clock_skew_seconds":0}
state is working (in a day, at the keyboard), paused (in a day, on a
break) or idle (the agent is running, the person is not in a day). at
carries the agent's own UTC offset. The server answers with the cadence it
wants rather than letting each agent pick one: report every
interval_seconds, and after stale_after_seconds of silence the pulse is no
longer believed.
A stamp more than a minute ahead of the server is refused with 400 rather
than accepted and corrected: a machine whose clock is wrong uploads hours that
are wrong too, and only its owner can fix that. clock_skew_seconds reports
the difference on every pulse so kasl can say so before it becomes a mystery.
The dashboard reads it back:
$ curl -H "Cookie: kasl_session=..." http://127.0.0.1:8080/api/v1/team/live
{"poll_seconds":30,"stale_after_seconds":180,
"members":[
{"user_id":"c49ea6a8-...","status":"working","since_received":12},
{"user_id":"0073460d-...","status":"offline","since_received":4210},
{"user_id":"9f1c2b40-...","status":"unknown","since_received":null}]}
offline and unknown are different answers. offline means a machine
was reporting and stopped; unknown means no pulse has ever arrived - no
agent, or a kasl too old to send one. Neither is reported as idle, because
the server does not know that, and a dashboard that guessed would tell a
manager their whole team stopped working the day they rolled out an older
agent. Where there is no pulse, the row falls back to what /team/days
knows: "day open", "last data 20 min ago", "never reported".
Its own endpoint, deliberately. The week's hours are a page load; this is
polled every poll_seconds while the tab is open, and folding the two together
would re-run the heaviest query on the server on a timer. It applies the same
visibility rule as the rest of the team endpoints - who is at their keyboard
this minute is more sensitive than a week's totals, not less - and the web UI
stops polling entirely while its tab is hidden.
Only the state is sent. Not the task, not the reason for the break: those belong to the day, under the privacy level that governs it. The pulse is listed in the privacy manifest at every level, as the latest claim only - replaced each time, never kept as a history. The reasoning is in ADR 0014.
Managing the team
Once an administrator exists, people and agent tokens are managed over the API rather than through the host's environment.
$ curl -X POST http://127.0.0.1:8080/api/v1/users -H "Cookie: kasl_session=..." -H "Content-Type: application/json" -d '{"email":"ivan@example.com","display_name":"Ivan","password":"..."}'
{"id":"9b5c1fd8-cf3d-433e-bb9e-0c2bf1c1cfac"}
$ curl -X POST http://127.0.0.1:8080/api/v1/users/9b5c1fd8-.../agents -H "Cookie: kasl_session=..." -H "Content-Type: application/json" -d '{"name":"ivan-laptop"}'
{"id":"b749b090-db08-464d-b48d-4fe15f7acc43","name":"ivan-laptop",
"token":"kasl_<64 hex chars>",
"notice":"this token is shown once; the server keeps only its hash"}
$ curl -X DELETE http://127.0.0.1:8080/api/v1/agents/b749b090-... -H "Cookie: kasl_session=..."
# 204; the same token now gets 401 from the ingest routes
| Route | Who |
|---|---|
GET /users, GET /users/{id}/agents |
admin (everyone), manager (their departments) |
GET /departments |
admin, manager |
POST /departments, PATCH/DELETE /departments/{id} |
admin |
PUT /users/{id}/department |
admin |
POST /users, PATCH /users/{id} |
admin |
POST /users/{id}/agents, DELETE /agents/{id} |
admin |
POST /auth/password |
anyone signed in, for their own password |
GET /audit |
admin |
A manager reads their departments and changes nothing. A department names its manager, and a person belongs to one:
$ curl -X POST http://127.0.0.1:8080/api/v1/departments -H "Cookie: kasl_session=..." -H "Content-Type: application/json" -d '{"name":"Engineering","manager_id":"d7c9ef3a-..."}'
{"id":"997c3947-4028-45e6-9c1c-9cd334b10c5d"}
$ curl -X PUT http://127.0.0.1:8080/api/v1/users/<id>/department -H "Cookie: kasl_session=..." -H "Content-Type: application/json" -d '{"department_id":"997c3947-..."}'
# 204; `{"department_id":null}` takes them out again without deleting anything
The manager of Engineering sees the people in Engineering, plus themselves — a manager who runs nothing yet would otherwise get an empty page and think the product was broken. An administrator sees everyone.
Someone with no department is visible to the administrator alone. Showing the unfiled to every manager, so nobody gets lost, fails in the direction nobody observes: forget to file a person and they are exposed company-wide, silently. Missing from a list is reported the same afternoon.
Deleting a department leaves its people unfiled rather than deleting them, and an employee cannot be made to run one — they could not see it, so it would silently have no working head. Issuing an agent token stays with the administrator: it is the authority to write someone's history, and there is no audit log until the next milestone.
An administrator sets an initial password and hands it over; the person
changes it with POST /auth/password, which requires the current one. The
server has no mail channel, so there is nothing to send an invite link to that
would not be handed over the same way a password is.
Some things follow from a change rather than being asked for separately:
- Deactivating someone, or resetting their password, deletes their sessions.
- Changing your own password ends every other session and keeps the one you are using.
- The last administrator cannot be demoted or deactivated — the only way back
from that is the
adminsubcommand on the host. - A user is never deleted, only deactivated: their days have to keep an owner.
Agent tokens are shown once and stored as a SHA-256. The reasoning is in ADR 0008 and ADR 0009.
The web UI
The same binary that answers the API serves the web app, on the same port:
http://127.0.0.1:8080 is the sign-in screen, and /api/v1/... is the API. A
self-hosted install is one file - there is no web server to configure, and no
way for the UI to be from a different build than the API it calls.
My week is the employee's own history: seven days, each drawn as a timeline of gold stretches broken by the pauses in them, and any day opens to its pauses and the tasks logged on it. The team is the manager's dashboard - a row per person with their hours, bars that compare people with each other, and a status that says what the server knows, under a band naming the people the server thinks are worth a look; clicking a row opens that person's twelve-week chart and their week, in the same component the personal page uses. The month is the same team as a grid - a square per person per day, shaded by hours, with weekends marked from the calendar and nothing recorded drawn as an empty square rather than as a worked day of zero. What is stored about me renders the manifest from what the server actually enforces (ADR 0011) rather than describing it again in the page, so the two cannot disagree.
The team screens are hidden from an employee's navigation, but that is tidiness rather than security: the endpoints behind them refuse an employee outright.
Where the installation's privacy level withheld something, the page says so in
that spot. A coarse day draws no timeline and states how many interruptions
there were and how long they came to, because an unbroken gold bar would be a
claim about the day that the server did not keep the evidence for.
The version in the header comes from /health - the server's, not the
bundle's. One product, one number.
Working on it
$ pnpm --dir frontend install
$ pnpm --dir frontend dev # http://localhost:5173, proxies /api to :8080
Vite serves the app and proxies /api to the server, so the browser sees one
origin and the session cookie behaves as it will in production. A change is a
page reload rather than a Rust compile.
$ pnpm --dir frontend lint # eslint, tsc and the unit tests
$ pnpm --dir frontend build # what gets embedded
TypeScript is held at 6.x on purpose. typescript-eslint does not run under
TS 7 yet, so upgrading takes linting with it - pnpm update --latest pulls 7
and pnpm lint then fails before it type-checks anything. The ^6.0.3 range
in frontend/package.json is what keeps that from happening by accident.
cargo build --release embeds whatever is in frontend/dist at that moment.
The repository carries an empty placeholder there, so a checkout without Node
still compiles - and a binary built that way says
no web UI was built into this binary instead of serving a blank page. The
decision and its trade-offs are in
ADR 0012.
What the server stores about you
An employee is asked to run an agent that notices when they stop typing. The honest answer to "what does it send" is one the server enforces and can recite, so it is an endpoint rather than a paragraph:
$ curl -H "Authorization: Bearer $KASL_TOKEN" http://127.0.0.1:8080/api/v1/privacy/agent
{"level":"full",
"summary":"This server stores your working hours, every interruption with the reason you gave for it, and the tasks you logged with their comments.",
"stored":[{"what":"workdays","detail":"the date, when the day started, and when it ended"},
{"what":"pauses","detail":"each interruption: when it began, how long it lasted, ..."},
{"what":"tasks","detail":"what you logged: the name, your comment, and how complete you marked it"},
{"what":"pause reasons","detail":"the text you type when you take a break by hand"},
{"what":"account","detail":"your email, display name, role, department, ..."},
{"what":"live status","detail":"whether your agent currently reports you as working, on a break, ..."}],
"never_collected":["keystrokes or what you type","window titles","which applications you run",
"screenshots or camera images","web pages you visit","file names or paths","your location"],
"visible_to":["you, in your own account","the manager of your department","administrators of this installation"],
"retention":"Kept for as long as the installation keeps it: there is no automatic deletion. ...",
"on_change":"Changing this setting affects what arrives from now on. ..."}
An agent reads it with its own token, so kasl can show it in the CLI - where the
employee already is - instead of requiring a login to the server that watches
them. Anyone signed in reads the same manifest at GET /api/v1/privacy.
Three levels
How much detail an installation keeps is a setting, and the server enforces it on the way in. A field a level excludes is dropped before the day is written: it never reaches the database, never reaches a backup, and no later change of mind can recover it.
| Level | Workdays | Pauses | Tasks | Free text |
|---|---|---|---|---|
full (default) |
hours | each one, timed | name and comment | kept |
moderate |
hours | each one, timed | name only | dropped |
coarse |
hours | how many, how long in total | not stored | dropped |
The live status sits outside this table on purpose. The levels govern what is stored about a day; the pulse is the agent saying what is happening right now, and the server keeps only the latest one - replaced each time it arrives, never accumulated into a record of when you were at your desk. It is named in the manifest at every level, because a manifest that listed only days would describe a quieter server than the one running.
Under coarse a day still records how much of it was paused, as a count and a
total. Without that the day would claim uninterrupted work, which is a more
flattering picture than the truth and a false one.
full is the default because the alternative is a breaking change disguised as
a virtue: a timid default would silently start discarding data in installations
already running. Narrowing is the deliberate act, and it is one request:
$ curl -X PUT -H "Cookie: kasl_session=..." -H "Content-Type: application/json" -d '{"level":"moderate"}' http://127.0.0.1:8080/api/v1/privacy
Only an administrator may set it, and the change goes into the audit log with both ends of it - a policy that can be quietly loosened is not a policy.
An upload says what it dropped. The response to POST /api/v1/days carries
the level that applied, and a discarded object when the level left something
out:
Told "5 pauses accepted" by a server that stored none, an agent would report a break as recorded when it was not.
Changing the level does not rewrite history. Narrowing it stops new detail from arriving and leaves what is already stored; widening it does not bring back what was dropped. A day the agent re-sends, though, is stored under the level in force now.
There is no per-employee opt-out. It was considered and rejected: a manager comparing a team where one person's pauses are missing gets a dashboard that lies by omission. The unit of the promise is the installation. The reasoning is in ADR 0011.
The audit log
Everything that changes people, departments or agent tokens is recorded, along with sign-ins and the attempts that failed:
$ curl -H "Cookie: kasl_session=..." "http://127.0.0.1:8080/api/v1/audit?limit=2"
[{"id":4,"actor_id":null,"actor_email":"ivan@example.com","action":"auth.login_failed",
"target_id":null,"target_label":null,"details":null,"at":"2026-08-22T00:58:12.880538Z"},
{"id":3,"actor_id":"85440341-...","actor_email":"boss@example.com","action":"agent.issued",
"target_id":"dcb60120-...","target_label":"ivan-laptop",
"details":{"user_id":"9dce4dd0-..."},"at":"2026-08-22T00:58:11.850690Z"}]
Filter with actor_id, target_id, action, since, until, and page with
limit (500 at most) and offset. "Everything that happened to this person" is
?target_id=....
Nothing secret goes in. An issued token is recorded as having been issued, never as a value; a password change is recorded as having happened. A failed sign-in keeps the address that was tried — a run of them against one account is the thing worth seeing — but never the password, which is often a real one belonging to somewhere else.
There is no route to delete from it. Not for old entries, not for a date range. A journal the watched party can erase is not a journal, and the administrator is the log'"'"'s main subject; trimming it is an operation for whoever holds the database. Reading the log is not itself recorded — an audit of the audit buries the actions under a log of people looking at the log.
Only an administrator may read it. The reasoning is in ADR 0010.
Importing history from before the server
Someone can track their time with kasl for a year before their team runs a server. That history is an ordinary SQLite file on their machine, and it does not have to be lost because the server arrived second:
$ kasl-server import --db kasl.db --user employee@example.com --timezone -03:00
read 240 workdays, 312 pauses, 460 tasks from kasl.db
skipped 17 tasks the employee had deleted
imported 240 days as employee@example.com at -03:00
--timezone is required and has no default. kasl stores bare wall-clock text,
so nothing in the file says which offset it was recorded in - and a wrong guess
produces a perfectly plausible-looking year of work at the wrong hour. The
answer comes from whoever knows, and is echoed back so it is on the record.
--dry-runreads and reports without writing anything.--since/--untilbound the import by date, both ends inclusive. This is how someone who moved between time zones is imported correctly: one run per stretch, each with the offset that stretch was recorded in.- The account must already exist - an import will not create it, so a typo in the email address cannot quietly file a year of history under a stranger.
- Re-importing replaces rather than duplicates, so a run that failed partway can simply be repeated, and a wrong offset is fixed by importing again with the right one.
The agent's file is opened read-only and never written to. Details and the trade-offs behind the fixed offset are in ADR 0006.
The data model
Migrations live in migrations/ and are applied on startup. The shape follows
kasl's own model, so a reader who knows the agent recognizes it:
| Table | Holds |
|---|---|
users |
People and their role: admin, manager, employee |
sessions |
Browser sign-ins; a token hash each, never the token |
departments |
Groups of people, each naming the manager who runs it |
audit_log |
Who did what, to whom, and when; append-only |
settings |
One row: how much detail this installation stores |
agents |
Installed kasl instances; a token hash each, never the token |
workdays |
One row per person per date: when the day started and ended, and - under a coarse privacy level - how much of it was paused |
pauses |
Idle stretches and manual breaks inside a day |
tasks, tags, task_tags |
What was worked on, and how it is labelled |
reports |
That a report was submitted, when, and with which figures |
Two differences from the agent's database are deliberate: instants are stored with a time zone (the agent stores bare wall-clock text, which does not survive a team spread across zones), and rows are tied together by foreign keys rather than by comparing dates. Both are recorded in ADR 0003.
Installing it
Docker and nothing else — no Rust, no Node, no build:
$ curl -o docker-compose.yml \
https://raw.githubusercontent.com/lacodda/kasl-server/main/docker-compose.install.yml
$ printf 'POSTGRES_PASSWORD=%s\n' "$(openssl rand -base64 24)" > .env
$ chmod 600 .env
$ docker compose up -d
$ docker compose logs server | grep -A 4 'administrator account'
An administrator account was created, because this installation had none:
email: admin@kasl.local
password: 6fp35gainu7zpj2yfwy3
Open http://localhost:8080, sign in with that, and change it. The password
is printed once and stored nowhere else — the alternative, writing one into a
file the server reads at every boot, leaves a credential lying around forever.
To name the administrator yourself instead, set KASL_ADMIN=email:password
before the first start.
Nothing to look at yet? KASL_DEMO=true docker compose up -d on an empty
database seeds a fictional team — see The demo.
The image is ghcr.io/lacodda/kasl-server, built for amd64 and arm64, so the
same compose file works on a laptop and on a Raspberry Pi. Pin a version in
production (KASL_VERSION=0.14.1); latest is for a first look.
Two more things before this holds a team's hours:
- Put it behind HTTPS and set
KASL_SECURE_COOKIES=true. Over plainhttp://the session cookie cannot carrySecure, which is why the compose file ships with it off — convenient for a trial, wrong for anything real. - Take backups. See below; the database's volume is the only copy until you do.
Building from source instead
docker-compose.prod.yml builds the image on the machine that will run it,
which is what this project's own stand does — it makes every deploy a proof
that the code compiles for that architecture. It takes about fifteen minutes on
a Pi 4 the first time; later builds reuse the cached layers.
$ docker compose -f docker-compose.prod.yml up -d --build
In both cases the database publishes no port — only the server reaches it, over the compose network — and the server runs as an unprivileged user.
Backups
The whole installation goes into one file and comes back out of it:
$ docker compose exec server kasl-server backup > kasl-$(date +%F).jsonl
wrote 4213 rows from 12 tables
$ docker compose exec -T server kasl-server restore < kasl-2026-08-28.jsonl
restored 4213 rows into 12 tables
The file is JSON Lines — one line per table — so it reads in any tool and
compresses well. --out and --from take paths instead of the standard
streams.
A restore refuses a database that already holds accounts. Merging two installations would mean deciding what wins, and every answer to that is wrong for somebody; restore into an empty database and the decision stays yours.
And it refuses a backup from a newer server. A file carries the schema version it was taken at, because the failure being avoided is not an error — it is a restore that appears to work while quietly dropping columns the older schema has no place for.
Agent tokens survive a restore, so the machines in the field keep reporting
without being re-enrolled — which matters most on the day you actually need
this. If you already have a backup regime, pg_dump remains available and this
does not replace it.
Configuration
Everything comes from the environment:
| Variable | Meaning | Default |
|---|---|---|
DATABASE_URL |
PostgreSQL connection string | required |
KASL_SERVER_ADDR |
Address the HTTP server binds to | 0.0.0.0:8080 |
KASL_AGENTS |
Agents to provision on startup, as email:token pairs separated by commas |
none |
KASL_ADMIN |
First administrator, as email:password. Unset, the server generates one on a first run |
none |
KASL_ADMIN_EMAIL |
Email for that generated administrator | admin@kasl.local |
KASL_SECURE_COOKIES |
Whether the session cookie carries Secure. Set false only when serving plain http:// |
true |
KASL_DEMO |
Seed a fictional team on an empty database, and refuse to start on one that holds real accounts. See The demo | false |
KASL_MAX_BATCH_DAYS |
Days one /days/batch request may carry |
31 |
KASL_MAX_BODY_BYTES |
Largest request body accepted | 4194304 |
RUST_LOG |
Log filter (tracing syntax) | kasl_server=info,tower_http=info |
Database migrations are embedded in the binary and applied on startup.
The privacy level is deliberately not here. It is set through the API and the change is audited; an operator editing a file and restarting leaves nothing behind that says who loosened the policy or when.
KASL_AGENTS is how the first agents get in while the admin UI does not exist
yet: each entry becomes an employee and an agent holding that token's hash.
Re-running with a changed token rotates it and revokes the old one. Tokens are
secrets — pass them through your deployment's secret store, not a committed
file — and the variable stops being the way in once tokens are issued from the
UI.
What it will do
- Ingest work-time data from kasl agents: workdays, pauses, tasks, reports (days, backfill and history import: done)
- Manager dashboards: who is working right now, hours per person, trends over time (the team's week, a drill-down into one person, the live status, the month as a heatmap, and signals with a twelve-week trend: done)
- Personal pages: every employee sees their own history (their own week, with the day timeline: done)
- Roles: admin, manager, employee (done)
- Self-hosted: a single binary — API and web UI in one file — plus PostgreSQL, so your data stays on your infrastructure (published image, install guide and backups: done)
- A demo: a fictional team on an empty database, to see the dashboards before installing anything (done)
Stack
Rust REST API (axum) + PostgreSQL (sqlx); the web UI is React 19 + TypeScript + Vite + Tailwind 4, built into the binary. Architectural decisions are recorded in docs/adr/.