Skip to content

Who can work. Who is out. Until when.

Your team's AI coding limits, hour by hour.

Tachly shows each person's 5-hour and weekly windows, who runs out before the reset, and how sure every number is. It runs on your own server.

Claude Code today. OpenAI Codex next — in the plan.

Example team · synthetic data

Simulated day · updated 14:20

Example team with synthetic data, at 14:20. Olena K.: ≈ runs out 15:40 · resets 18:00; Taras: ≈ 44% left · no status line; Iryna: 84% left · resets 15:00; Max: Weekly limit hit · back Thursday 09:00; Dana: 86% left · resets 18:00; Yurii: 92% left · resets 19:00.

Example team, synthetic data: each person's limit by hour at 14:20. Hours up to 14:20 are as recorded, later hours are the forecast.
Person00:0001:0002:0003:0004:0005:0006:0007:0008:0009:0010:0011:0012:0013:0014:0015:0016:0017:0018:0019:0020:0021:0022:0023:00
Olena K.capacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftlimit hitlimit hitcapacity leftcapacity left≈ capacity left (estimate)≈ limit hit (forecast)≈ limit hit (forecast)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)
Taras≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)
Irynacapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity left≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)
Maxlimit hitlimit hitlimit hitlimit hitlimit hitlimit hitlimit hitlimit hitlimit hitlimit hitlimit hitlimit hitlimit hitlimit hitlimit hitlimit hitlimit hitlimit hitlimit hitlimit hitlimit hitlimit hitlimit hitlimit hit
Danano datano datano datano datano datano datano datano data≥ at least, part of the window missing≥ at least, part of the window missing≥ at least, part of the window missing≥ at least, part of the window missingcapacity leftcapacity leftcapacity left≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)
Yuriicapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity leftcapacity left≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)≈ capacity left (estimate)
  • Capacity left
  • Limit hit, out until reset
  • ≈ Estimate or forecast
  • ≈ Out, forecast
  • ≥ At least
  • No data
  • Now
  • Window resets

Every cell says how it knows

A number you can't trust is worse than none. So every figure carries its source, in the cell and in the words next to it.

  • Measured

    Read from the agent's own limit readout on that person's machine: the status line in a terminal; for the Claude desktop app, the plugin asks the signed-in terminal claude.

  • ≈ Estimate

    No status line on that machine, or a forecast to the reset: worked out from tokens and the team's calibration.

  • ≥ At least

    Part of the window is missing, say history imported later. The real figure is this or higher.

  • No data

    Nothing arrived for that hour. It stays a gap and is never drawn as zero.

Red means two things only: now, and the moment a window resets.

  • Now
  • Window resets

Open a row. See where the limit went.

The same views open for every person; members see only their own. Shown here: Olena K. from the example team, synthetic data.

Limits burndown

Used so far and forecast to the reset, for the 5-hour and the weekly window: who runs out before the window resets, and by how much.

≈ 2 h 20 min without the agent before the 18:00 reset

Week71% used · ≈ 88% by Monday's reset

Tokens and cache

Input, output, cache reads and cache writes, hour by hour, so a session that keeps re-reading its context stands out.

935K tokens so far today · 91% read from cache

Cost against budget

At API prices, per person and for the team, against the month's budget, with a forecast to the month's end.

$412 of $600 this month · ≈ $585 by month end

Cost this month by person, example team
Olena K.$148
Dana$87
Taras$74
Iryna$53
Max$41
Yurii$9

Coach

Plain checks on the habits that burn limits, each with the fix rather than a score.

Expensive models on simple tasks

Opus/Fable on short tasks: 5 of 16 sessions had < 2K output tokens and no subagents (≈$6.40 in these sessions). The matching Sonnet costs about 2× less per output token.

Olena K. · this week · synthetic

Also, per person

  • Sessions with their subagents, skills, plugins and MCP tools
  • Search across prompt text, where it is collected
  • Projects and repositories, with cost and time for each

From each machine to your server. Nowhere else.

The Tachly plugin on each person's machine reads the agent's hooks, its session transcripts and, once turned on, its status line. It sends what it reads to the server you run.

  1. Each machine

    The coding agent with the Tachly plugin. Offline, it keeps a queue and sends it later.

  2. Your server

    The Tachly web app behind HTTPS. Accounts with two-factor sign-in; Google sign-in for invited addresses.

  3. Your databases

    Postgres keeps accounts, people, devices and settings.

    ClickHouse keeps sessions, messages, tokens and limit readouts.

Sent

  • When sessions run: timestamps, model, project folder and branch
  • Token counts: input, output, cache reads and writes
  • Names of the tools, subagents, skills and MCP servers used
  • The repository's origin URL (credentials stripped), the machine's hostname and OS
  • Session titles you set, with secrets masked
  • Limit readouts, once the status line is on: /tachly:statusline on
  • Prompt text, only while collection is on for that person: on by default, and the admin can turn it off per person. Secrets such as API keys are masked first; with collection off, only the prompt's length and the name of a slash command are sent.

Never sent

  • The agent's replies and its reasoning
  • Tool inputs and outputs, and the contents of your files
  • Anything to a third party: there is no hosted service

The invite says what is collected before anyone connects a machine, and the admin can turn prompt text off for each person.

Joining a team?

Your admin sends you a personal invite. Three steps, and you are on the schedule.

  1. Sign in

    Open the invite and set a password, or continue with Google if your admin turned it on and invited that exact address. Two-factor sign-in is there for everyone and required for admins.

  2. Connect your machine

    In the coding agent, add the plugin and sign in. The browser opens to confirm. Once per machine.

    /plugin marketplace add tachly/tachly-plugin
    /plugin install tachly@tachly
    /tachly:login https://tachly.dev:3000
  3. Check what is collected

    Shows the server, the send queue and whether your prompt text is collected.

    /tachly:status
Your admin sees
Your sessions, tokens, cost, limits and tools, your repositories and machine names. Your prompt text only while collection is on for you.
Your admin doesn't see
The agent's replies, tool output or the contents of your files. Tachly never sends them.
You see
Your own data only: your limits, sessions, cost and coach. Not your teammates'.

Run it on your own server

One Docker Compose stack: the web app, Postgres, ClickHouse and Caddy for HTTPS. Migrations run on every start.

You need a Linux VPS with Docker 24+ and Compose v2, at least 2 GB of RAM (4 GB for a larger team or a long history), and a domain pointed at it.

Contact us

Essentials in .env

TACHLY_DOMAIN
Your domain. Caddy gets and renews its certificate.
AUTH_SECRET
At least 32 bytes: openssl rand -base64 48
POSTGRES_PASSWORD
A long random password for the database
ADMIN_EMAIL · ADMIN_PASSWORD_HASH
The first admin, for the first start only; the bcrypt hash goes in single quotes
PLUGIN_REPO_URL
Optional: plugin source for the install lines. Defaults to the public plugin repository
GOOGLE_CLIENT_ID · GOOGLE_CLIENT_SECRET
Optional: Google sign-in for invited addresses

The full guide is docs/deploy.md in the repository.

Put your team's limits on the schedule.

Self-hosted. Your team's data stays on your server.