llm-cost-dashboard

Real-time terminal dashboard for LLM token spend -- cost per request, per-model breakdown, projected monthly bills, and budget enforcement. Zero external services required.

Built with ratatui and crossterm. Structured logging via tracing. No database. No network. No cloud account.

What it does

llm-dash reads a stream of LLM request records (JSON lines) from a log file or stdin, computes USD cost using a built-in pricing table, and renders a live terminal dashboard showing:

• Total spend (session) and projected monthly bill (extrapolated from the last hour of activity)
• Cost by model -- horizontal bar chart sorted by highest spend
• Recent requests -- scrollable table with timestamp, model, token counts, cost, and latency
• Budget gauge -- visual progress bar with configurable alert threshold and hard limit
• Spend sparkline -- mini chart of the last 60 request costs

Supported providers and models

Unknown models automatically fall back to $5.00/$15.00 input/output pricing. Lookup is case-insensitive.

Quickstart

Install from crates.io

cargo install llm-cost-dashboard

Install from source

git clone https://github.com/Mattbusel/llm-cost-dashboard
cd llm-cost-dashboard
cargo install --path .

Run with demo data

llm-dash --demo

Set a monthly budget and tail a log file

llm-dash --budget 50.0 --log-file requests.log

Pipe from your application

your-llm-app | llm-dash --budget 25.0

Dashboard layout

 LLM Cost Dashboard  [q: quit | r: reset | d: demo data | j/k: scroll]
+------------------+--------------------------------------------------+
| Summary          |  Cost by Model (uUSD)                            |
| Total: $0.0142   |  ████████ claude-sonnet-4-6                      |
| Proj:  $0.42/mo  |  ████ gpt-4o-mini                                |
+------------------+  ██ claude-haiku-4-5                             |
| Budget           +--------------------------------------------------+
| ████░░░ 14.2%    |  Recent Requests                                 |
| $8.58 remaining  |  12:34:01  claude-sonnet  847in/312out  $0.0031  |
+------------------+--------------------------------------------------+
| Sparkline: spend over last 60 requests                              |
+--------------------------------------------------------------------+

Log file format

Records must be newline-delimited JSON (NDJSON). The four required fields are model, input_tokens, output_tokens, and latency_ms:

{"model":"claude-sonnet-4-6","input_tokens":512,"output_tokens":256,"latency_ms":340}
{"model":"gpt-4o-mini","input_tokens":128,"output_tokens":64,"latency_ms":12}

Optional fields:

Malformed lines are skipped and logged as warnings -- the dashboard never crashes on bad input.

CLI reference

llm-dash [OPTIONS]

Options:
  --budget <BUDGET>        Monthly budget limit in USD [default: 10.0]
  --log-file <LOG_FILE>    JSON log file to tail for live data
  --demo                   Start with built-in demo data pre-loaded
  -h, --help               Print help
  -V, --version            Print version

Keyboard controls

Environment variables

Tracing output is written to stderr so it does not interfere with piped stdin/stdout.

Library usage

The crate exposes its core types as a library crate (llm_cost_dashboard) for embedding cost tracking directly in your application:

use llm_cost_dashboard::{CostLedger, CostRecord};

let mut ledger = CostLedger::new();
let record = CostRecord::new("gpt-4o-mini", "openai", 512, 256, 34);
ledger.add(record).expect("valid record");
println!("total: ${:.6}", ledger.total_usd());
println!("projected/mo: ${:.2}", ledger.projected_monthly_usd(1));

Key types

Architecture

src/
  main.rs          # CLI entry point (clap + tracing init)
  lib.rs           # Public re-exports
  error.rs         # DashboardError (thiserror)
  cost/
    mod.rs         # CostRecord, CostLedger, ModelStats
    pricing.rs     # Static pricing table + lookup/compute_cost
  budget/
    mod.rs         # BudgetEnvelope (hard limit + alert threshold)
  log/
    mod.rs         # LogEntry, RequestLog, IncomingRecord (NDJSON parser)
  trace/
    mod.rs         # TraceSpan, SpanStore (distributed tracing helpers)
  ui/
    mod.rs         # App state + run() event loop
    dashboard.rs   # Full-frame layout compositor
    widgets.rs     # Budget gauge, sparkline, summary panel
    theme.rs       # Centralised colour/style palette

tests/
  unit_tests.rs        # Public-API unit tests (pricing, ledger, budget, log)
  integration_tests.rs # Cross-module integration tests
  integration.rs       # End-to-end app-level tests

benches/
  cost_bench.rs    # Criterion benchmarks for pricing lookup and aggregation

Development

# Run all tests
cargo test

# Run with debug tracing
RUST_LOG=debug cargo run -- --demo

# Lint
cargo clippy --all-targets --all-features -- -D warnings

# Format
cargo fmt

# Benchmarks
cargo bench

# Documentation
cargo doc --open

Related projects by @Mattbusel

• tokio-prompt-orchestrator -- Rust async LLM pipeline orchestration
• rot-signals-api -- Options signal REST API
• prompt-observatory -- LLM interpretability dashboard

License

MIT -- see LICENSE for details.