Read your Claude Code history in the browser

Drop a session file from ~/.claude/projects. Every prompt and reply, with each tool call folded under the step that made it and every compaction marked. Search across sessions, download one as Markdown or HTML, or keep the ones worth keeping as pages. Nothing is uploaded: the file is read on your computer and never leaves it.

Drop your session files here

The .jsonl files in ~/.claude/projects, one per session. One file, or a project folder's worth at once.

Read on your computer. Nothing is uploaded.

Finding the files, and what to do with them

01
Find the files
Claude Code writes every session to ~/.claude/projects: one folder per project, named after its path with the slashes turned into hyphens, and inside it one .jsonl file per session, named by the session's id. On Windows that is C:\Users\you\.claude\projects. A session that ran subagents also has a folder of its own beside its file, with their files under subagents.
02
Drop one, or a folder's worth
Select every .jsonl in a project folder at once and each becomes a session in the list, newest first, with a search box across all of them. A subagent's file is listed too, and says that it is one.
03
Read it the way it ran
Your prompts, Claude's replies, and under each step the tool calls it made, folded: the command, the path or the query on the line, and what came back inside. Where the context was compacted, a marker says so with the token counts before and after.
04
Take it out, or keep it
One session as Markdown with a line per tool call, or as one HTML file that is this view with the calls still folded; all of them as a zip; or as pages in a vault, the prompts and replies with the tool calls counted, which is the record worth searching next month.

What it reads

Every prompt you typed, and every slash command, as what was typed
Every reply, with one that arrived in several records read as one step
Every tool call, with its input where it reads (a command, an edit as a diff, a path and its lines) and the first 2,000 characters of what came back, the rest counted
Where the context was compacted, with the token counts before and after
Claude Code's own running cost and line counts, where the file has them
Files from a session's subagents folder, marked as subagent sessions

What it leaves out

Thinking blocks. They are counted on the session, not shown.
Images: a screenshot you pasted, or one a tool took. Counted, never kept.
Context the program injected: CLAUDE.md, skills, reminders, the system prompt. The file records them as attachments and meta records, and they are not your words.
Bookkeeping: modes, file snapshots and request records, which are most of the lines in a file.

Questions

Is my history uploaded anywhere?: No. A session file holds everything the agent saw, which can include your code and the output of every command it ran, so this page reads it on your computer and the sessions stay in this tab until you close it. The only button that sends anything is Import, which puts the sessions you chose into your own Sumibako vault and nowhere else. The browser's network panel will show nothing leaving while you read, search or download.
Where does Claude Code keep its history?: Under ~/.claude/projects in your home folder, which on Windows is C:\Users\you\.claude\projects. Each project you have run it in has a folder named after the project's path with the slashes turned into hyphens, and each session in it is one .jsonl file named by the session's id. Starting Claude Code again in the same folder starts a new file; nothing you did in the terminal deletes an old one.
What does the compaction marker mean?: When a session's context fills up, Claude Code summarises what came before and carries on from the summary; the marker is where that happened, with how many tokens the context held before and after. It is also the moment anything you had not written down was at risk, which is what the post linked below is about.
Why does a session read shorter than it felt?: Three things are left out on purpose. Thinking is counted, not shown. A tool's result is cut at 2,000 characters, with the rest counted. And the context the program injected, such as CLAUDE.md, a skill or a reminder, is not shown at all, because it is not something you said. A session that was mostly the agent reading files will read short.
How is this different from Claude Code's own export?: The /export command in Claude Code writes the conversation as text, the way it looked in the terminal. The .jsonl file is the full record underneath it, with every tool call, its input and result, the compactions and the cost, which is what this reads. And it works for sessions you can no longer open, as long as the file is still there.
How big a file can it read?: A day's work is often a file of 20 to 40 MB, and it reads in a few seconds. Everything stays in the tab's memory, so a very large set of files at once is the one thing that can be slow; drop a project folder at a time rather than every project you have.
What does importing into Sumibako put there?: One page per session under a parent page, holding the prompts and replies with each run of tool calls counted on one line, so a long session fits. The free plan holds up to 1,000 notes; if the sessions are more than the room you have, the import stops at the limit and says how many did not fit. The whole import can be undone in one step from Settings.

What a compaction costs you, and the habit that makes it cost nothing, is in the post on Claude Code compaction. Where the rest of what Claude Code writes ends up is in Where Claude Code saves plans.

Keep the sessions worth keeping

A .jsonl in a hidden folder is where a week of decisions goes to be forgotten. A page in a vault is searchable from any device, can be shared with a link, and is where the agent can write the next handoff itself.

Start writing - free How an agent writes into a vault

Read your Claude Code history in the browser

Finding the files, and what to do with them

Find the files

Claude Code writes every session to ~/.claude/projects: one folder per project, named after its path with the slashes turned into hyphens, and inside it one .jsonl file per session, named by the session's id. On Windows that is C:\Users\you\.claude\projects. A session that ran subagents also has a folder of its own beside its file, with their files under subagents.

Drop one, or a folder's worth

Select every .jsonl in a project folder at once and each becomes a session in the list, newest first, with a search box across all of them. A subagent's file is listed too, and says that it is one.

Read it the way it ran

Your prompts, Claude's replies, and under each step the tool calls it made, folded: the command, the path or the query on the line, and what came back inside. Where the context was compacted, a marker says so with the token counts before and after.

Take it out, or keep it

One session as Markdown with a line per tool call, or as one HTML file that is this view with the calls still folded; all of them as a zip; or as pages in a vault, the prompts and replies with the tool calls counted, which is the record worth searching next month.

What it reads

Every prompt you typed, and every slash command, as what was typed

Every reply, with one that arrived in several records read as one step

Every tool call, with its input where it reads (a command, an edit as a diff, a path and its lines) and the first 2,000 characters of what came back, the rest counted

Where the context was compacted, with the token counts before and after

Claude Code's own running cost and line counts, where the file has them

Files from a session's subagents folder, marked as subagent sessions

What it leaves out

Thinking blocks. They are counted on the session, not shown.

Images: a screenshot you pasted, or one a tool took. Counted, never kept.

Context the program injected: CLAUDE.md, skills, reminders, the system prompt. The file records them as attachments and meta records, and they are not your words.

Bookkeeping: modes, file snapshots and request records, which are most of the lines in a file.

Questions

Is my history uploaded anywhere?

No. A session file holds everything the agent saw, which can include your code and the output of every command it ran, so this page reads it on your computer and the sessions stay in this tab until you close it. The only button that sends anything is Import, which puts the sessions you chose into your own Sumibako vault and nowhere else. The browser's network panel will show nothing leaving while you read, search or download.

Where does Claude Code keep its history?

Under ~/.claude/projects in your home folder, which on Windows is C:\Users\you\.claude\projects. Each project you have run it in has a folder named after the project's path with the slashes turned into hyphens, and each session in it is one .jsonl file named by the session's id. Starting Claude Code again in the same folder starts a new file; nothing you did in the terminal deletes an old one.

What does the compaction marker mean?

When a session's context fills up, Claude Code summarises what came before and carries on from the summary; the marker is where that happened, with how many tokens the context held before and after. It is also the moment anything you had not written down was at risk, which is what the post linked below is about.

Why does a session read shorter than it felt?

Three things are left out on purpose. Thinking is counted, not shown. A tool's result is cut at 2,000 characters, with the rest counted. And the context the program injected, such as CLAUDE.md, a skill or a reminder, is not shown at all, because it is not something you said. A session that was mostly the agent reading files will read short.

How is this different from Claude Code's own export?

The /export command in Claude Code writes the conversation as text, the way it looked in the terminal. The .jsonl file is the full record underneath it, with every tool call, its input and result, the compactions and the cost, which is what this reads. And it works for sessions you can no longer open, as long as the file is still there.

How big a file can it read?

A day's work is often a file of 20 to 40 MB, and it reads in a few seconds. Everything stays in the tab's memory, so a very large set of files at once is the one thing that can be slow; drop a project folder at a time rather than every project you have.

What does importing into Sumibako put there?

One page per session under a parent page, holding the prompts and replies with each run of tool calls counted on one line, so a long session fits. The free plan holds up to 1,000 notes; if the sessions are more than the room you have, the import stops at the limit and says how many did not fit. The whole import can be undone in one step from Settings.

What a compaction costs you, and the habit that makes it cost nothing, is in the post on Claude Code compaction. Where the rest of what Claude Code writes ends up is in Where Claude Code saves plans.