Why LangCost?

You run OpenClaw and your LLM bill keeps climbing — but you can't tell why.

The provider dashboard shows total tokens. It doesn't show that your agent looped 12 times on the same tool call, or that 30% of output tokens were unnecessarily verbose, or that prompt caching wasn't working.

LangCost reads your session logs and tells you exactly what's wasting money:

Scanned 12 sessions from openclaw
├── 12 traces, 1,812 spans, 1,948 messages
├── Total cost: $73.24
├── Estimated waste: $8.60 (11.7%)
└── Top waste: high_output (72.7%), tool_failure_waste (26.7%), low_cache (0.6%)

No API keys. No cloud. Everything runs locally. Your data never leaves your machine.

Quick Start

Three commands. That's it.

# Install the CLI + the OpenClaw adapter
npm install -g langcost @langcost/adapter-openclaw
# or: bun add -g langcost @langcost/adapter-openclaw

# Scan your sessions
langcost scan --source openclaw

# Open the dashboard
langcost dashboard

Why two packages? langcost is the core CLI — analysis engine, dashboard, and reports. Adapters are plugins that read data from specific sources. You install only the adapters you need. Right now OpenClaw is the only adapter — more are coming.

LangCost auto-detects your OpenClaw installation at ~/.openclaw, ingests your sessions, runs waste analysis, and serves a local dashboard at http://localhost:3737.

# Point to a custom OpenClaw directory
langcost scan --source openclaw --path /path/to/openclaw

# Analyze a single session file
langcost scan --source openclaw --file /path/to/session.jsonl

# Scan older sessions (default is last 30 days)
langcost scan --source openclaw --since 90d

# Force re-analysis of everything
langcost scan --source openclaw --force

Features

🔍 Waste Detection

Six rules that automatically find wasted spend in every session:

| | Rule | What it finds | |:---:|------|-------------| | 🔴 | Tool Failures | Failed tool calls that burned tokens for nothing — bash: command not found, 12 calls failed, $1.65 wasted | | 🟡 | Agent Loops | Agent stuck calling the same tools in a cycle — read → bash → read → bash repeated 8 times | | 🟡 | Retry Patterns | User re-prompting because the agent failed — 3 similar messages in a row, agent struggling | | 🟠 | High Output | Spans with output 3x+ the session average — one response used 4,200 tokens when peers averaged 380 | | 🟢 | Low Cache | Prompt caching disabled or underused — paying full input price on every call | | 🔵 | Model Insight | Flags expensive model usage — 100% Opus usage, helps you decide when cheaper models suffice |

Every finding includes the dollar amount wasted and a specific recommendation to fix it.

🔬 Trace Explorer

Expand any session to see the full execution timeline — every LLM call and tool call in order:

#1   LLM   opus-4-5   in:2.8K  out:141   $0.017   ok
 ├── read   README.md                                ok
 └── read   src/main.ts                              ok
#2   LLM   opus-4-5   in:5.2K  out:380   $0.083   ok
 ├── bash   ls -la                                   ok
 ├── write  src/fix.ts                               ok
 └── bash   bun test                                 ✗ error
#3   LLM   opus-4-5   in:8.1K  out:520   $0.130   ok
 └── edit   src/fix.ts                               ok

Read exactly what the agent did, which tools it called, what failed, and what each step cost.

📊 Dashboard

A local web UI at localhost:3737:

• Trace table — all sessions with cost, waste, status. Sortable and filterable.
• Expandable rows — click to see waste findings + execution timeline inline
• Cost overview — total spend, waste percentage, cost-over-time chart
• Model insights — which models you're using and what they cost
• Recommendations — prioritized list of what to fix first

💻 CLI Reports

# All sessions sorted by cost
langcost report --sort cost

Trace                    │ Model    │   Cost │  Waste │ Status
─────────────────────────┼──────────┼────────┼────────┼────────
before-compaction        │ opus-4-5 │ $42.60 │  $6.03 │ error
expensive-session        │ opus-4   │  $0.28 │  $0.11 │ ok
simple-session           │ sonnet-4 │ $0.002 │ $0.001 │ ok

# Deep dive into one session
langcost report --trace <trace-id>

# Only sessions with tool failures
langcost report --category tool_failure_waste

# JSON for scripting
langcost report --format json

💰 22 Models Supported

Built-in pricing for Anthropic, OpenAI, Google, DeepSeek, and Mistral:

| Provider | Models | |----------|--------| | Anthropic | Opus 4, Sonnet 4, Haiku 4.5, Haiku 3.5 | | OpenAI | GPT-4.1, GPT-4.1-mini, GPT-4.1-nano, GPT-4o, GPT-4o-mini, o3, o3-mini, o4-mini | | Google | Gemini 2.5 Pro, 2.5 Flash, 2.0 Flash, 2.0 Flash Lite | | DeepSeek | V3 (chat), R1 (reasoner) | | Mistral | Large, Small, Codestral |

Using a self-hosted or unlisted model? Costs show as $0 but all token counts and waste detection still work. Custom pricing support is coming soon.

How It Works

  ~/.openclaw/sessions/*.jsonl
         │
         ▼
    ┌─────────┐
    │  ingest  │  Read JSONL → normalize to traces, spans, messages
    └────┬────┘
         ▼
    ┌─────────┐
    │ analyze  │  Run 6 waste detection rules
    └────┬────┘
         ▼
    ┌─────────┐
    │  store   │  SQLite at ~/.langcost/langcost.db
    └────┬────┘
         │
    ┌────┴────┐
    ▼         ▼
  CLI      Dashboard
 report   localhost:3737

• Everything runs locally — no cloud, no API keys, no tracking
• Data stays in a single SQLite file on your machine
• Keeps the 500 most recent sessions to manage disk space
• Plugin architecture — adapters handle data ingestion, analyzers handle intelligence

CLI Reference

langcost scan --source <adapter> [options]
  --source <adapter>      Required. "openclaw"
  --path <path>           Override data source path
  --file <path>           Analyze a single session file
  --since <duration>      Default: 30d. Accepts: 7d, 30d, 90d, all
  --force                 Re-ingest and re-analyze everything
  --db <path>             Override database path

langcost report [options]
  --format <fmt>          table (default) | json | markdown
  --sort <field>          cost | waste | date
  --limit <n>             Number of traces (default: 20)
  --trace <id>            Detailed single-trace report
  --category <cat>        Filter by waste category
  --db <path>             Override database path

langcost dashboard [options]
  --port <port>           Default: 3737
  --db <path>             Override database path

langcost status
  --db <path>             Override database path

Upcoming

| | Feature | Description | |:---:|---------|-------------| | 🧭 | Fault Attribution | Trace failures backwards to find the root cause — not just which step errored, but which upstream agent caused it | | 🧩 | More Waste Rules | Unused tool schemas, duplicate RAG chunks, unbounded conversation history, uncached system prompts | | 🔌 | More Adapters | Pluggable data sources beyond OpenClaw — bring your own traces | | 🏷️ | Custom Model Pricing | Set input/output/cache prices for self-hosted and unlisted models |

Contributing

LangCost has a plugin architecture. Three easy ways to contribute:

🧩 Add a waste rule — standalone function in packages/analyzers/src/rules/. Copy an existing rule as a starting point.

💲 Update model pricing — edit packages/core/src/pricing/providers.ts. Add new models or fix outdated prices.

🔌 Build an adapter — npm package implementing IAdapter from @langcost/core. The CLI discovers it automatically.

Tech Stack