slop.boutique

A lab for computation and intelligence as material and medium. Run by Manuel and Ramon.

032 · 23 October 2026 · agents · developer-tools

Transcript Hub

A shared hub for Claude Code transcripts: upload sessions, read them turn by turn, annotate as a team and query them from JavaScript.

Previewin your browser
Three cards from a session on a grey background: a failed npm test run with its error output and AI failure note, the note's thread with a reply, and an Edit call's diff that chains saves on a queue

In early access for subscribers until 22 November 2026. Subscribe to run it now.

Subscribe →

A shared hub for Claude Code transcripts. Drop the JSONL session files from ~/.claude/projects/, and the team can read every session turn by turn, annotate what went wrong or right, and query the whole collection from JavaScript.

  • Upload in the browser: files are parsed locally, likely secrets (keys, tokens, private keys) are redacted, folders can be skipped by name, and you review the list before anything is stored. Subagent files nest under the tool call that started them.
  • Read sessions: prompts, thinking, tool calls with their input, output, exit codes and diffs; corrections and interrupts are flagged. Each session shows duration, active time, prompts, tool calls, failures, tokens and cost.
  • Annotate as a team: notes on a session, a turn or a single tool call, in categories such as ai-failure, environment-issue, success or to-improve, with tags, replies and resolve.
  • Markers, files and collections: jump between errors, interrupts, corrections, compactions and subagents; see which files a session read and wrote; group sessions into collections.
  • Query and analyse: a query builder over sessions and tool calls with parameters, saved queries, results as a page, a JS console, an API reference, and analytics by model, repo and uploader with a list of where agents struggled.
  • An agent in the page: Claude can answer questions about the transcripts and act on the page by running hub code you can read step by step.

It starts empty: the hub holds only the transcripts your team uploads. The screenshots use made-up sessions.

The full hub: sessions table with uploader, start, duration, prompts, tools, failures, output tokens and cost, and the agent panel with suggested questions
The full hub: sessions table with uploader, start, duration, prompts, tools, failures, output tokens and cost, and the agent panel with suggested questions
A session about rate limiting, read turn by turn: a failed pip install, an interrupt and a correction from the user, then the new token bucket file written in full
A session about rate limiting, read turn by turn: a failed pip install, an interrupt and a correction from the user, then the new token bucket file written in full
The Queries page: the failing-sessions query builder with a minFailures parameter, Run, Save, Explain and Open as page, and its result table
The Queries page: the failing-sessions query builder with a minFailures parameter, Run, Save, Explain and Open as page, and its result table
Close-up of an Edit call: the diff that wraps disk.write in a promise queue, and its output
Close-up of an Edit call: the diff that wraps disk.write in a promise queue, and its output
The Markers tab: counts for errors, interrupts, corrections, compactions, subagents, edits and notes, with the two errors listed and Prev and Next buttons
The Markers tab: counts for errors, interrupts, corrections, compactions, subagents, edits and notes, with the two errors listed and Prev and Next buttons
Close-up of a failed Bash call: npm test exits 1 with sync keeps order expected [a,b,c] got [a,c,b], and a team note marks it as an AI failure
Close-up of a failed Bash call: npm test exits 1 with sync keeps order expected [a,b,c] got [a,c,b], and a team note marks it as an AI failure
The JS console in the side panel: hub.sessions().select('title', 'toolFailures').orderBy('-toolFailures').limit(5) and its JSON result
The JS console in the side panel: hub.sessions().select('title', 'toolFailures').orderBy('-toolFailures').limit(5) and its JSON result
Close-up of the sessions table: titles with repo, branch, model and collection, uploaders, durations, failures in red, output-token meters, cost and note flags
Close-up of the sessions table: titles with repo, branch, model and collection, uploaders, durations, failures in red, output-token meters, cost and note flags
Analytics across ten sessions: totals for prompts, tool calls, fail rate, corrections, active time, tokens and cost, tables by model, repo and uploader, and where agents struggled
Analytics across ten sessions: totals for prompts, tool calls, fail rate, corrections, active time, tokens and cost, tables by model, repo and uploader, and where agents struggled
Close-up of a Write call: a dependency-free token bucket class in server/limits.py shown as new contents
Close-up of a Write call: a dependency-free token bucket class in server/limits.py shown as new contents
The API tab: the hub reference listing builders, helpers and fields available to queries, the console and the agent
The API tab: the hub reference listing builders, helpers and fields available to queries, the console and the agent