Skip to main content

​
Cost Management & Budget Control

Myrm provides a comprehensive multi-layer cost control system that ensures you never overspend on LLM usage — while maintaining task quality through intelligent routing.

​
Budget Policy

​
Configuration (Settings → System → Budget Policy)

Set spending limits across three independent dimensions: Each dimension triggers a 4-level progressive response:

​
How It Works

  1. WARNING — Agent receives a budget hint via cache-safe prompt injection. Eco mode activates (compression threshold lowered 20%)
  2. FINALIZATION — Agent is instructed to output final results immediately. Tool calls are stripped from model output
  3. EXCEEDED — Hard block. No further LLM calls until budget is reset or expanded
Budget status is displayed in real-time via SSE push to the frontend BudgetBadge and Goal status card.

​
Real-Time Cost Observability

​
Per-Message Display

Every assistant message shows detailed token breakdown:
  • 7 token types: prompt, completion, cached, cache_write, reasoning, citation, total
  • Cost in USD with cache savings percentage
  • Model tier indicator (simple / standard / reasoning)
  • Cache break reason attribution when cache misses occur

​
Goal Panel (Active Goals)

While a goal is running, the expanded status card shows:
  • Burn rate — tokens/minute consumption
  • Cost rate — $/minute spending
  • ETA — estimated time to budget exhaustion
  • Progress bar — percentage of budget consumed

​
Usage Statistics Dashboard (Settings → System)

Full analytics dashboard with:
  • Daily spending chart (7/30/365 day ranges)
  • Session-level drill-down with token/cost breakdown
  • Model breakdown panel (cost per model)
  • Hour and weekday distribution heatmaps
  • Per-agent usage cards with sparkline trends
  • Routing analytics (how many requests used each tier)

​
Full-Chain Accounting — No Blind Spots

Every LLM call, streaming or not, lands in the same token ledger:
  • Unified recording — streaming and non-streaming calls are counted in one place (the adapter layer), so sub-agent spawns, dynamic workflows, background summaries, memory extraction, context compression and wiki compilation are all attributed correctly
  • Honest failure paths — a failed LLM call never fabricates usage, so the ledger never inflates
  • End-of-run accuracy — message_end events are back-filled with real usage from the tracker, so per-run totals match the sum of individual calls
For hosted / metered-billing users this means the bill you see is the bill you can trust — background work is never hidden, and nothing is double-counted.

​
Intelligent Cost Routing

​
ComplexityRouter (Automatic)

Myrm automatically routes requests to the most cost-effective model tier:
  • SIMPLE — Quick questions, greetings → cheapest model
  • STANDARD — Normal tasks → balanced model
  • REASONING — Complex multi-step problems → premium model
Features:
  • Two-phase classification (rule scoring + LLM judge)
  • Session momentum — follow-up messages inherit the conversation’s recent tier to prevent quality degradation
  • Penalty feedback — records misroutes and adjusts future routing probability (24h decay)
  • Min-tier floor — regenerate requests automatically escalate tier
  • Zero configuration — active by default on all accounts

​
Privacy-Aware Routing

Sensitive content is automatically routed to privacy-appropriate models based on PII detection level.

​
Subagent Cost Isolation

​
Delegation Budget

Each root task has a max_descendants = 20 guard preventing runaway subagent spawning. This ensures a single task cannot create an unbounded delegation tree that drains budget.

​
Per-Root Budget Scope

Budget tracking is scoped per root agent run — subagent costs accumulate against the parent’s budget dimensions, providing unified visibility.

​
Multi-Agent Cost Pre-Flight

When the agent needs to spawn multiple sub-agents (batch delegation, council cross-review, or alternatives parallel generation), Myrm estimates the total LLM cost before execution begins:
  • Batch mode — estimates cost based on task count × per-task model pricing
  • Council mode — accounts for expert_count × (1 + cross_review_rounds) + 1 (chair synthesis) LLM calls
  • Alternatives mode — estimates based on the number of parallel expert agents
If the estimated cost exceeds $0.50, an interactive approval card appears in the chat, showing:
  • Total estimated cost in USD
  • Number of LLM calls to be made
  • Remaining budget
You can approve to proceed or reject to cancel the operation — no tokens are consumed until you approve. This prevents surprise bills from complex multi-agent operations.

​
Long Task Resilience

​
Checkpoint & Recovery

If the server process is interrupted during a long task:
  1. Signal-safe checkpoint — all running subagents are checkpointed to disk (JSON + fcntl file locking)
  2. Orphan recovery — on restart, orphan scanner detects interrupted checkpoints and publishes events to the UI
  3. Resume API — user can resume from checkpoint with full state restoration (messages + workspace)

​
Goal Continuation Guard

The continuation engine evaluates 9+ conditions before each turn:
  • Budget remaining check
  • Cancellation / steering token
  • Convergence detection
  • Goal drift detection (trajectory scoring every 5 turns)
  • Per-todo checkpoint (optional pause after each step)
  • Protected file integrity validation

​
Output Compression

CLI tool output is automatically compressed using intelligent rules to minimize token consumption:
  • Path deduplication and summarization
  • Repeated pattern collapsing
  • Binary/noise filtering
  • Large output truncation with summary

​
Stable Prefix for Cache Efficiency

System prompts and tool schemas are carefully ordered and canonicalized to maximize prompt cache hit rates:
  • Deterministic tool schema ordering (SSOT)
  • Schema canonicalization (removing non-semantic differences)
  • Cache breakpoint detection with attribution reporting

​
Test Coverage

313 tests verify the cost management system: