Reported issues for AI Usage

Pod holds 7 of 7 GitHub reports that passed its relevance review. This can include external user reports, maintainer-confirmed bugs, and concrete feature gaps. Treat them as evidence to inspect, not a count of distinct defects.

Back to AI Usage.

Most discussed

Install fails with node-gyp errors before the user can read the fix

Problem

better-sqlite3 is a native addon. On Windows with npm 10 the install dies inside node-gyp — npm 10 there ignores the dependency's gypfile: false flag and compiles from source even though a usable prebuilt binary is bundled, and that fails without Visual Studio Build Tools.

The reason this is the worst problem in the project is where it fails: during npx. The README has a troubleshooting entry explaining the one-line fix, but the user never gets that far — they see a…

Read the thread · 2026-09-02 · closed · 0 comments

MCP clients prompt for permission on every tool call

Problem

All seven MCP tools (usage_summary, session_usage, model_usage, client_usage, recent_sessions, project_usage, daily_usage) are pure readers: they read the collectors' source files and our own SQLite, and nothing else.

None of them declares that. Without tool annotations, a client has no way to know a tool is safe, so it falls back to its default and raises a confirmation prompt on every call.…

Read the thread · 2026-09-02 · closed · 0 comments

A stale MCP server never tells anyone, only the CLI does

Problem

#17 shipped a staleness check in 0.3.0 and closed with a known limit spelled out in its own description:

This only reaches people who run the CLI. A global install driven purely through Claude Code still sees nothing.

That limit is the whole population the feature was for. The reasoning at the time was that npx -y ai-usage-mcp re-resolves on a cold start, so server users are fine. Two cases it misses:

npm i -g ai-usage-mcp wired into an MCP config. Pinned until…

Read the thread · 2026-09-01 · closed · 0 comments

Budget guard via MCP elicitation

The one thing a batch CLI structurally cannot do. When a tool call finds spend over a declared threshold, don't return text — raise a form elicitation. Claude Code renders a real dialog ("$18.40 today against a $15 cap — continue / raise cap / stop") and the answer comes back to the server.

This answers ccusage discussion #358 ("Ability for CC to check its own usage%?"), which their maintainer never answered, and ccusage #483 (real quota limits), which they declined.

Two constraints that…

Read the thread · 2026-09-01 · open · 0 comments

Add MCP prompts and resources

The server exposes plain tools only — no resources, no prompts. Both are confirmed supported by Claude Code as an MCP client, and both are free UI surface:

Prompts appear as slash commands: /ai-usage:daily-review, /ai-usage:why-was-today-expensive
Resources appear as @-mentions: @ai-usage:usage://today pulls cost into context with no tool call at all

Section 10 of PROJECT_CONTEXT.md deferred these to "later". Later has arrived.

Scope

Two or three prompts…

Read the thread · 2026-09-01 · closed · 0 comments

Expose daily usage as an MCP tool, and move the CLI onto the formatter

UsageService.dailyUsage() (src/services/usage-service.ts:178) is public and reachable from the MCP layer, but no tool calls it — the per-day breakdown is CLI-only.

Separately, the daily CLI command is the one place a frontend formats output inline (src/cli/index.ts:137-147) instead of going through the formatter, which breaks the layering rule the rest of the codebase follows.

Scope

formatDaily in src/services/formatter.ts
Move the CLI daily command onto it,…

Read the thread · 2026-09-01 · closed · 0 comments

Surface per-project usage (the code exists but is unreachable)

project_path is collected by both collectors, stored, and indexed — and then never surfaced. UsageRepository.byProject() (src/db/repositories/usage-repository.ts:246) is fully implemented and has zero call sites; the projectPath filter (:110) and UsageQuery.projectPath (src/services/usage-service.ts:27) exist too. The capability is written but unreachable from any interface.

"Which repository is my spend going to" is a question no competing tool answers, and on a real machine the…

Read the thread · 2026-09-01 · closed · 0 comments

Most recent

The remaining reports are on the project's issue tracker.