Skip to content
runmarq

Features

Estimates

  • A plain answer to whether to wait: almost done, a few minutes, time for something else, or a long task. It leads while the evidence is thin; the likely range leads once it is not.
  • A likely range for the time left, and how long it could take when the tail runs long. Never a bare countdown.
  • “Finishing up” when the work checks out and the agent stops starting tools, once your own runs show those signs mean the end.
  • Low, medium or high confidence, from the evidence behind the estimate.
  • A named reason whenever the estimate moves, such as repeated test failures.
  • The estimate pauses while the agent waits for your approval.
  • A run that outlasts similar runs gets a longer estimate instead of reaching zero.
  • Messages you send mid-task and background-task notices continue the same run, so one task gets one estimate.
  • Plan progress from Claude's task list or Codex's plan, when the agent keeps one.
  • “Not enough information yet” rather than an invented number.

Estimate

About 6–11 min

Time for something else

Elapsed
4 min
Most likely
About 9 min more
Prediction confidence
Medium

Learning

  • Four models on this computer. Once about 30 runs have finished, each is tested on your recent runs as if they were live and counts by how well it did.
  • A history baseline from the first run, starting from deliberately wide built-in values.
  • A phase simulation once ten runs with tool activity have finished.
  • Similar earlier prompts, compared through salted word vectors.
  • A gradient-boosted model after 100 finished runs, used only if it beats the others.
  • Ranges calibrated to contain about 80% of outcomes once about 34 runs have finished.
  • Retraining in the background as runs finish, without slowing the app.

Estimation model

  • History baseline64 runsIn use
  • Phase simulation61 runsIn use
  • Similar prompts64 runsIn use
  • Learned model64 of 100 runsNot in use

Signals

  • Model responses, output and thinking tokens, context in use and compactions, from Claude Code and Codex transcripts.
  • Files and lines changed in the working tree during the run, from git.
  • Learned durations for the commands your agents run, such as your test suite.
  • The effort level and model the agent reports.
  • The current phase: planning, exploring, editing, testing and building, running commands, delegating or compacting context.
  • The numbers only, never the text. Token counts, working-tree changes and prompt comparison can each be turned off in Settings → Estimation signals.

Run activity

Effort
High
Phase
Editing
Model responses
35 + 6 by subagents
Output tokens
16.4k (2.1k thinking)
Context in use
84k of 200k
Compactions
1
Lines edited
118
Working tree
4 files, 132 lines changed

Tray and alerts

  • A tray menu with every active run and its estimate.
  • A compact panel for the corner of your screen, optionally kept on top.
  • Notifications when a run finishes, fails, needs approval or stops reporting, with an optional sound.
  • A warning when a run looks stuck: the same failure again and again, a command running far longer than usual, or a run long past its estimate with nothing happening.
  • Quiet mode, and no finish or failure alerts for runs shorter than you choose.
  • Clear run states: running, waiting for you, finished, failed, cancelled and status unknown.
  • Starts when you log in, if you want it to.

Claude Code finished

Run in acme-web · 7 min

Notifications

  • Quiet mode
  • When a run finishes
  • When a run fails
  • When an agent needs approval
  • When an agent stops reporting
  • When a run looks stuck
  • Play a sound
  • Skip very short runs20 seconds

Usage

  • Output tokens, responses, runs and working time per agent, for today, the last 7 days and the current 5-hour window.
  • When the current window started and when it resets.
  • An optional budget per window that you set yourself, shown as progress.
  • Counted from the transcripts on this computer. Not your plan's official meter: input and cached tokens are not counted, and runmarq does not know your plan's limits.

Claude Code output tokens

5-hour window
126k
Today
212k
Last 7 days
1.4M
126k / 200k

Terminal

  • runmarq wait blocks until the run finishes and exits 0 when it did, so runmarq wait && npm run deploy runs only after a clean finish. Supporting terminals show its progress on the tab and taskbar.
  • runmarq statusline prints one short line for a shell prompt, tmux or Claude Code's status line, and prints nothing when the collector is not running.

$ runmarq wait && npm run deploy

Waiting for Claude Code · acme-web · 4:12 elapsed · A few minutes, about 2–5 min

✓ Claude Code · acme-web finished after 6:12.

$ runmarq statusline

runmarq ● 2 working · next: about 3–7 min

History and insights

  • History with search and filters by agent, status, repository and date.
  • For each run: phases, events, active and waiting time, and every estimate change with its reason.
  • A forecast chart that sets each estimate against the actual duration.
  • Time spent waiting for approvals, and typical durations by agent and task type.
  • Prediction quality on your machine: how often the range held, median error, and error by forecast age and agent.
  • CSV export of your timing data.

Phases

Run finished · 7 min

  • Reading files56 s · 6 calls
  • Editing code1 min · 4 calls
  • Running tests3 min · 3 calls · 2 failed
  • Waiting for you42 s
  • Editing code38 s · 2 calls
  • Running tests34 s · 1 call

Integrations

  • Claude Code, Codex and Cursor, each through its official hooks.
  • Detects which agents are installed on this computer.
  • Shows the exact file and entries before writing, and keeps a backup.
  • Connects every project, or one project only.
  • Marks an integration connected only once its first real event arrives.
  • Removal takes out only the entries runmarq added.
  • runmarq run -- codex closes Codex runs cleanly when the process exits.

Claude Code

Detected · 2.1.143
Approve and connect

Claude Code

Connected
Re-run setup

Command line

Privacy

  • No account, no analytics and no network requests to compute an estimate.
  • None of your agents' model usage is spent on estimates.
  • Prompts are classified inside the hook process and discarded. Short titles are stored only if you turn them on.
  • Source code, file contents, command arguments and output, and transcript text are never stored.
  • Repository folder names can be hidden. History can be deleted at any time, or automatically after 30, 90 or 365 days.
Privacy statement

What it does not do

  • It knows when an agent's turn ended, not whether your goal was met.
  • It does not read terminal output or anything on your screen.
  • Cursor's approval prompts are not visible to it, and agent sessions started without hooks cannot be observed.
  • Estimates are wide until your own runs have been recorded.
Integration guides

Try it on your next run.

Build from sourceGetting started