aitools

A one-row bar above the prompt showing how close your Claude subscription is to its limits, and what the work would cost at Anthropic API list prices.

AI Tools    5h 42% ·1h12m    wk 18% ·Thu    ctx 31% 62k/200k    cache 42m    thread $3.12 (+$0.41)    month $184.20    limit ~3:40pm   [🛠] [⁝]

| Segment | Meaning | |---|---| | AI Tools | A dim label naming the bar | | 5h 42% ·1h12m | 5-hour usage window: percent used, time until it resets | | wk 18% ·Thu | Weekly window: percent used, the day it resets (hours when under a day) | | ctx 31% 62k/200k | Context window fill: percent and tokens used / window size | | cache 42m | How long the prompt cache should stay warm: Claude Code writes the main conversation's cache with a 1-hour TTL, restarted by every request. Amber under 10 minutes, cache cold once expired. An estimate from request times; the API does not report cache state, and Anthropic may evict early | | thread $3.12 | This session's cost at API prices | | (+$0.41) | What the last turn cost (last $0.41 when thread cost is hidden) | | month $184.20 | This calendar month's Claude Code usage at API prices, across all sessions on this machine | | limit ~3:40pm | At your pace over the last hour, when the 5-hour window would run out; shown only if before it resets |

Usage and context turn yellow at 70% and red at 90%. A value not known yet shows —; until the first response of a session the bar shows the figures from your previous session.

Show or hide the bar

/aitools toggles the whole bar (buttons included); /aitools on and /aitools off set it. The choice holds across sessions. /handoff works either way.

Settings

Press ⁝ to open the options above the bar, five to a row, and toggle any of: the AI Tools label, 5-hour usage, weekly usage, reset countdowns, context %, context tokens, cache warmth, thread cost, last-turn cost, monthly cost, burn-rate projection, threshold colors, threshold alerts, Tools menu. Changes show in the bar right away and are saved across sessions. Press ⁝ again to close it; opening it closes the 🛠 menu, and the other way round.

Threshold alerts are one-time toasts when the 5-hour or weekly window passes 90%, and when context passes 80%.

Handoff

/handoff asks the current conversation for a brief (goal, current state, decisions, open tasks, key files, next step), prints it in the chat as formatted text, and copies the raw markdown to the clipboard. Paste it into a new session to continue there. The brief is a command output row, so the current session's model reads it too.

The 🛠 button opens a row above the bar with Handoff, Workflows, Task View, Clean View and Agent Dock, right-aligned; press it again to close the row without picking. Handoff writes the same brief into a Handoff brief pane, drawn as formatted text with a Copy button; nothing is copied until you press it, and the brief stays out of the chat (so the current session's model does not read it). Workflows runs Claude Code's built-in /workflows directly where the engine has it (the terminal); in the desktop Code tab, which answers a typed /workflows itself, it puts /workflows in the prompt box for you to send with Enter.

Task View and Clean View

Two views from the 🛠 menu that turn Claude's work into a checklist under the bar (an open menu still shows above it): done steps marked with a green ✓, the step in progress with a blue ●, steps not started with a dim ○. In the menu a ● marks the view that is on and a ○ the one that is off; picking the one that is on turns it off, and picking the other switches. The choice holds across sessions.

While one is on, its name shows in the status line under the prompt (with Agent Dock after a comma while the dock is open). /taskview and /cleanview do the same as the menu from the prompt: no argument toggles that view (turning it on switches off the other), on turns it on and off turns it off.

Adding dark mode                                                    3 of 5
✓ Read the theme code                                   100% ▰▰▰▰▰▰▰▰▰▰
✓ Add the color tokens                                  100% ▰▰▰▰▰▰▰▰▰▰
● Wire up the toggle                                     40% ▰▰▰▰▱▱▱▱▱▱
○ Update tests                                            0% ▱▱▱▱▱▱▱▱▱▱
○ Run checks                                              0% ▱▱▱▱▱▱▱▱▱▱

Each step has a bar: full once done, empty until it starts, and for the step in progress Claude's own estimate of how far along it is, which moves each time Claude updates the checklist (the note asks for about every quarter of a step).

• Task View shows the checklist and leaves the chat as it is (only the checklist tool's own rows are hidden).
• Clean View shows the checklist and hides every tool call and everything Claude writes while it works; only the final reply of each finished request stays. It hides tool calls and in-progress replies already in the chat too.

While a view is on, each prompt you send carries a hidden note asking Claude to plan the work as plain-language steps and report them through the mod's checklist tool (mcp__aitools__checklist), updating it as each step starts and finishes. The mod answers that tool itself, before Claude Code's permission check, so it should not ask for permission. Each new prompt starts a fresh checklist; once every step is done and Claude stops, it folds to one line, ✓ Done: <title> (<n> steps). A quick question that needs no work gets no checklist.

Clean View recognizes a final reply as the last thing Claude wrote before your next prompt. It reads them from the conversation when the mod loads (so replies from before a reload or restart still show) and adds each new one as a request finishes.

Agent Dock

/agentdock, or Agent Dock in the 🛠 menu (● while open), opens the Agent Dock pane for handing a big job to a team of helper agents (extra copies of Claude working in parallel); either again, or the pane's own close mark, closes it. It stays open until you close it, and comes back after the mod reloads. While the dock is closed, or its pane is not on screen (the status line then reads Agent Dock (not shown)), nothing below applies.

Team size: ○ 1 ○ 3 ● 5 ○ 10 ○ 20 ○ 30   Helpers: ● Fast & cheap ○ Same as chat

5 agents · 3 working · 2 idle · 1 queued · 1 done
● Research business licenses · Reading the city site               40% ▰▰▰▰▱▱▱▱▱▱  1:20
✓ Draft the floor plan                                             100% ▰▰▰▰▰▰▰▰▰▰  0:56

• Team: how many helpers may run at once (1, 3, 5, 10, 20 or 30). Past 10 the dock warns in yellow that a big team uses your Claude usage much faster.
• Helpers: Fast & cheap runs every helper request on Sonnet 5.5 at low reasoning; Same as chat uses the chat's model and reasoning level.
• While the dock is open, each prompt carries a hidden note asking Claude to split a job with separate parts across the team, starting the pieces in parallel. Each helper's prompt asks it to report what it is doing and how far along it is through the mod's agent_progress tool (mcp__aitools__agent_progress), which fills its card.
• A full team refuses further starts with a note to try again when a helper finishes (a hook may not hold a start for long); those pieces count as queued until they start.
• Cards show the task, what the helper is doing, its percent (its own estimate), a bar and its running time; a running helper's text is a mid grey, finished ones get a green ✓ and failed ones a red ✕. When every helper is done and Claude has replied, a line under the cards says how long they took (All 3 helpers finished in 6 to 10 seconds each.). The cards and that line stay until your next prompt.
• The team and helper picks are saved; the dock itself starts closed in a new session.

How monthly cost is computed

The mod reads Claude Code's own transcripts (~/.claude/projects/**/*.jsonl, or under CLAUDE_CONFIG_DIR), counts each assistant response once (by message id) when its timestamp falls in the current month, and prices its token usage (input, output, cache writes at the 5-minute or 1-hour rate, cache reads) with the table in src/pricing.ts (API list prices as of 2026-09-25). Only bytes added since the last scan are read; the cache lives in the mod's store and resets each month. A ~ before a figure means part of it used an estimated rate (an unknown model priced by its family) or a file could not be read.

Limits

• Claude Code only: the Code tab in Claude Desktop and the terminal. Conversations in Claude Desktop's chat tab are not visible to mods and are not counted.
• Usage percentages come from the API's rate-limit headers, so they update after each response.
• Dollar figures are what the same tokens would cost on the API, not what the subscription costs.