Claude Code Decoder

“How much did my Claude Code session actually cost?” Paste the .jsonl from ~/.claude/projects/<hash>/<session>.jsonl (or drag-drop the file). Get a full breakdown: per-model, per-tool, cumulative timeline, cache savings, insights.

Every byte parses in your browser. The page makes zero network calls on paste — the URL stays clean, your transcript stays private. Verify in the Network tab.

Decoder

Sample session is pre-loaded — click ✕ Clear sample and paste your own when ready.

🔒Your transcript stays in your browser. Parsing and cost math happen client-side via JavaScript. Nothing is uploaded to runlocalai.co or any third party. You can verify with the Network tab.

§ Paste or upload a Claude Code session

{"type":"user","message":{"role":"user","content":[{"type":"text","text":"Refactor src/api/auth.ts to use async/await and add tests."}]},"timestamp":"2026-05-14T10:00:00Z","uuid":"a1"}
{"type":"assistant","message":{"role":"assistant","model":"claude-sonnet-4-5","content":[{"type":"text","text":"I'll read the current implementation first."},{"type":"tool_use","id":"t1","name":"Read","input":{"file_path":"src/api/auth.ts"}}],"usage":{"input_tokens":2400,"output_tokens":85,"cache_creation_input_tokens":15000,"cache_read_input_tokens":0}},"timestamp":"2026-05-14T10:00:03Z","uuid":"a2"}
{"type":"user","message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"t1","content":"function login(email, password) {\n  return fetch('/api/login')...\n}"}]},"timestamp":"2026-05-14T10:00:04Z","uuid":"a3"}
{"type":"assistant","message":{"role":"assistant","model":"claude-sonnet-4-5","content":[{"type":"text","text":"Let me also check the existing tests."},{"type":"tool_use","id":"t2","name":"Glob","input":{"pattern":"src/api/**/*.test.ts"}}],"usage":{"input_tokens":350,"output_tokens":40,"cache_creation_input_tokens":0,"cache_read_input_tokens":17400}},"timestamp":"2026-05-14T10:00:06Z","uuid":"a4"}
{"type":"user","message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"t2","content":"src/api/auth.test.ts\nsrc/api/users.test.ts"}]},"timestamp":"2026-05-14T10:00:07Z","uuid":"a5"}
{"type":"assistant","message":{"role":"assistant","model":"claude-sonnet-4-5","content":[{"type":"tool_use","id":"t3","name":"Read","input":{"file_path":"src/api/auth.test.ts"}}],"usage":{"input_tokens":280,"output_tokens":30,"cache_creation_input_tokens":0,"cache_read_input_tokens":17750}},"timestamp":"2026-05-14T10:00:09Z","uuid":"a6"}
{"type":"user","message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"t3","content":"existing test content..."}]},"timestamp":"2026-05-14T10:00:10Z","uuid":"a7"}
{"type":"assistant","message":{"role":"assistant","model":"claude-sonnet-4-5","content":[{"type":"text","text":"Now I'll refactor."},{"type":"tool_use","id":"t4","name":"Edit","input":{"file_path":"src/api/auth.ts","old_string":"function login","new_string":"async function login"}}],"usage":{"input_tokens":420,"output_tokens":180,"cache_creation_input_tokens":0,"cache_read_input_tokens":18030}},"timestamp":"2026-05-14T10:00:15Z","uuid":"a8"}
{"type":"user","message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"t4","content":"Edit succeeded"}]},"timestamp":"2026-05-14T10:00:16Z","uuid":"a9"}
{"type":"assistant","message":{"role":"assistant","model":"claude-sonnet-4-5","content":[{"type":"text","text":"Running the tests to verify."},{"type":"tool_use","id":"t5","name":"Bash","input":{"command":"npm test src/api/auth.test.ts"}}],"usage":{"input_tokens":250,"output_tokens":50,"cache_creation_input_tokens":0,"cache_read_input_tokens":18250}},"timestamp":"2026-05-14T10:00:20Z","uuid":"a10"}
{"type":"user","message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"t5","content":"All 12 tests passed."}]},"timestamp":"2026-05-14T10:00:35Z","uuid":"a11"}
{"type":"assistant","message":{"role":"assistant","model":"claude-sonnet-4-5","content":[{"type":"text","text":"All tests pass. The refactor is complete — login now uses async/await and the existing test coverage still passes."}],"usage":{"input_tokens":280,"output_tokens":45,"cache_creation_input_tokens":0,"cache_read_input_tokens":18500}},"timestamp":"2026-05-14T10:00:37Z","uuid":"a12"}

12 events parsed1 user turns · 6 assistant turns · 5 tool calls

Session cost

$0.102across 1 model · 37s session

Input tokens

3,980

$0.012

Output tokens

430

$0.0065

Cache writes

15,000

$0.056

Cache reads

89,930

$0.027

§ Insights

✓Prompt caching saved approximately $0.24 on this session — Anthropic's cache-read rate is 10% of normal input, and it kicked in cleanly here.

ℹEdit · Write dominated this session (~3% of spend). That's typical for refactor / large-edit sessions — flagged so you can see where the budget went, not as a critique.

§ Cost by tool category

Edit · Write

$0.00(47% · 1 calls)

Read · Glob · Grep

$0.00(40% · 3 calls)

Bash · PowerShell

$0.00(13% · 1 calls)

Output cost is attributed to tool calls by proportional share of the turn's output tokens. Categories don't include input cost — input is shared across the whole turn and can't be cleanly split per tool.

§ Cost accumulation timeline

cumulative spend over session

claude-sonnet-4-5

§ Per-model breakdown

Model	Input	Output	Cache W	Cache R	Total	Share
Claude Sonnet 4.5	$0.012	$0.0065	$0.056	$0.027	$0.102	100%

§ Per-tool breakdown

Tool	Calls	~ Output tokens	Category
Read	2	115	read
Glob	1	40	read
Edit	1	180	write
Bash	1	50	execute

How to find your sessions

Claude Code stores every conversation as a JSON-lines file at the path below. Each line is one event (user message, assistant turn with usage, tool result, etc.).

# macOS / Linux
ls ~/.claude/projects/

# Each directory under projects/ is a project (hash of the cwd).
# Inside, every session is a separate .jsonl file:

ls ~/.claude/projects/<hash>/
# → 01HX0Z1ABCDE.jsonl
# → 01HX0Z2FGHIJ.jsonl
# → ...

# Pick the session you want to analyze, cat its contents, paste here.
cat ~/.claude/projects/<hash>/01HX0Z1ABCDE.jsonl

# On Windows: %USERPROFILE%\.claude\projects\

The decoder accepts any subset of a session — partial paste works fine for spot-checking. If you upload a file the entire JSONL is parsed at once.

Where to go from here

Cost vs Cloud →

Now that you know what your session cost on Claude, paste any prompt + see what it'd cost across 11 providers and your local rig.

Claude Code → local →

Step-by-step guide to swap Claude Sonnet/Opus for a local model in your Claude Code workflow. Real commands.

Stream Visualizer →

Watch tokens stream at your hardware's speed, side-by-side with Claude. The visceral version of the math you just saw.

Stack Builder →

If the bill above stung — eight inputs → full rig + runtime + model recipe to swap out Claude entirely.