💡 Live Cost Meter & Avatar-Spoken Budget Guardrails (BYOK Transparency)

[github-actions[bot]]

Summary

Surface real-time token spend, per-tool cost breakdowns, and a hard budget cap directly in the avatar UI — with the avatar verbally narrating cost milestones ("That last research turn cost about 12 cents — we're at $0.43 of your $2 session budget"). BYOK users currently fly blind: they pay Anthropic directly but have no in-app visibility into what each agentic loop is actually costing them, which is the #1 friction point for non-technical users adopting agentic workflows.

Market Signal

The Claude Agent SDK now ships a max_budget_usd parameter that hard-caps spend per session, plus first-class cost/usage tracking APIs (Python v0.1.48 / TS v0.2.71, March 2026). Anthropic explicitly documents cost tracking as a primary SDK concern. Meanwhile, every BYOK competitor (Cursor, Continue, Windsurf) is being criticized in dev forums for opaque billing — Cursor shifted to a usage-based pricing model in 2025 specifically because users demanded cost transparency. None of these tools surface cost in a way Sarah (PM/analyst persona) can act on mid-session.

User Signal

No direct user feedback exists yet (project pre-launch), but the PRD's Sarah persona is non-technical and explicitly cost-sensitive — a single runaway agentic loop on Opus could cost more than her monthly SaaS budget. The lack of an in-product cost ceiling is a known adoption blocker for BYOK in non-technical segments. Existing Discussion #55 (Effort Modes) hints at this concern but addresses model selection, not live spend visibility.

Technical Opportunity

The Claude Agent SDK exposes token usage on every message and supports max_budget_usd natively. The AgentBackend port in src/main/agent/ is the natural seam — wrap the SDK adapter to emit cost domain events through the existing DomainEventBus, persist them in the audit repository, and bridge them to the renderer via the agent:* IPC namespace. The avatar's existing TTS pipeline (renderer/speech/) already supports interrupt-driven utterances, so cost narration can ride the same channel as agent responses without a new subsystem. A simple BudgetPolicy value object in shared/types/domain/ keeps the limit logic testable.

Assessment

Dimension	Score	Rationale
Feasibility	high	SDK provides the primitives; only adapter, IPC, and a small UI surface needed. ~1 small epic.
Impact	high	Removes the single biggest BYOK adoption blocker for the target persona.
Urgency	high	Competitors are racing to fix BYOK opacity; SDK feature is fresh and not yet widely adopted.

Adversarial Review

Strongest objection: "Cost narration is annoying noise — users will mute it within a day, and the cost meter is just a number that no one looks at."

Rebuttal: Narration should be opt-in and budget-threshold-driven (only speaks at 50%/80%/100%), not a constant nag. The real value is the hard cap — max_budget_usd makes it impossible for an agent loop to silently burn $50, which is the actual fear. The visible meter is secondary; the budget guardrail is the load-bearing feature. This also unlocks per-workflow ROI conversations Sarah can take to her manager ("this saved me 3 hours and cost $0.40").

Suggested Next Step

Create a product brief for "BYOK Cost Transparency v1": scope the BudgetPolicy domain object, the agent:cost-update IPC event, the threshold-based avatar narration, and a minimal cost panel in the overlay. Spike on the SDK's max_budget_usd behavior under streaming to confirm partial-message cost emission.

---

Proposed by the BMAD Analyst (Mary) on 2026-04-07.
Workflow run

[github-actions[bot]]

Weekly Update

What Changed

Claude Cowork usage limits are now the #1 complaint in user reviews, providing strong market validation for this idea. Key new signals since this idea was created (April 7):

1. Claude Cowork launched and expanded: Available on macOS (Feb 2026), Windows (Feb 2026), and expanded to web/iOS/Android on July 7. Cloud remote sessions launched in beta (Max plan only, $200/month).

2. Usage limit frustration is documented across multiple review sites:

   • Pro users ($20/month) hit limits during intensive Cowork sessions
   • Even Max plan users ($200/month) report "rationing usage"
   • Anthropic's rolling 5-hour usage windows make limits unpredictable
   • "Race Mode" (parallel agent execution) is gated to $100/month Max plan
   • Sources: Substack (Intelligence & Power), LowCode Agency, VibeCoding, CoworkGuru

3. June 2026 billing change: Non-interactive use (GitHub Actions, headless agents, Agent SDK) now draws from dedicated monthly credits that don't roll over. This makes BYOK cost visibility structurally important.

4. Claude Agent SDK now includes per-agent cost attribution — making granular cost tracking technically feasible without manual token counting.

5. ChatGPT Voice Desktop launched July 23 requiring Plus/Pro/Business subscriptions. Microsoft Copilot requires M365 licensing. Every major competitor uses subscription-gated models with opaque limits.

Updated Assessment

Dimension	Previous	Current	Delta
Feasibility	high	high	→ SDK per-agent cost attribution makes implementation cleaner
Impact	high	high	→ Market validation confirmed — cost transparency is a top user need
Urgency	med	high	↑ Competitors launched with opaque limits; BYOK positioning window is now

Recommendation

Advance — promote to architectural requirement. The cost tracking data pipeline (extracting usage metadata from each API response) should be designed into the AgentMessage pipeline from Phase 1. The UI surface can be phased, but the data capture must be day-one. A related new discussion (#411) proposes the full BYOK Cost Intelligence Dashboard with competitive positioning framing.

---

Sources: Claude Cowork Review (Intelligence & Power), Claude Cowork Review (LowCode), Claude Cowork Limits (VibeCoding), ChatGPT Voice Launch (Fortune)