Orienting eats the meter.
Reads, searches, web and MCP retrieval, plus read-only shell like rg and git diff. Demo archive: 68% of estimated cost.
Dosu ships Decant · August 12
Decant reads the session logs Claude Code and Codex already write and turns them into numbers. On its demo archive — 119 sessions, $1,252.23 in estimated spend — 68% of the cost was context gathering before any file changed. The code that shipped was 3%.
Disclosure: Decant's launch post was written by Attention Deficit co-host Taylor Dolezal at Dosu. We are covering it anyway, on the record. Costs are estimated at standard API rates; subsidized plans will read differently than your bill.
WAL + FTS5 · full-text search across messages and tool calls
outbound calls. Decant ingests files already on your disk; sessions, data, and outputs stay on your machine.
One session, four buckets
Decant sorts every session into four buckets. Scrub through them — the shares shown are from the launch post's demo archive, not your setup.
Reads, searches, web and MCP retrieval, plus read-only shell like rg and git diff. Demo archive: 68% of estimated cost.
Reasoning blocks and plan tools — the model working out an approach before touching anything.
Structured edits plus the shell commands that write, build, or test files. Demo archive: 3% of estimated cost.
Replies, summaries, explanations — the visible part of the run.
Claude Code and Codex log every session — prompts, responses, tool calls, files touched — as JSONL on your disk. Decant normalizes both formats into a single SQLite archive with full-text search, then prices and indexes your history as it lands. Watch mode follows along with native filesystem events plus a periodic sweep.
agent sessions indexed in the archive Dosu used for the launch walkthrough.
total estimated cost across those sessions, priced at standard API rates.
No account, no telemetry, no outbound calls. The archive, the web app, and the API all run on your machine.
Decant prices sessions at standard API rates. If you're on a subscription or subsidized plan, the numbers show what the work would cost via the API — not your invoice.
Dosu calls that stretch orienting: the agent working out where things live and what's relevant. In the demo archive it dwarfs everything else — and it's exactly the work agents repeat run after run.
Shares are from the launch post's demo archive, not a benchmark. The post states the context and code figures; the middle bar is the remainder, which combines planning and communicating without an individual split. Decant's methodology doc describes how runs are categorized.
The CLI mirrors the app — decant stats, decant files — every read command takes --json, and a documented local API means your agents can query your own history. Session pages show the full transcript, the four-bucket split, the context-window chart with every compaction, and subagents included.
Recurring command history gets turned into reusable artifacts, with secrets redacted on the way out.
Suggested optimizations persist with implemented-state tracking, so the tool knows which advice you already took.
npx @dosu/decant@latest — opens the web app, syncs in the backgroundbrew tap dosu-ai/dosu then brew install decantbun run dev — Apache 2.0, TypeScript on Bun, local React UIThe logs were always on your disk. Now they answer questions.