Show HN: Seedeep – I couldn't see what Claude Code was doing, so I drew it

Seedeep, a new open-source tool by developer duqaXxX, visualizes Claude Code's internal operations in real time, showing every model call, subagent, and token cost during a session. Analyzing 1,830 real transcripts, it found that 39 of 47 failed API calls were the last event in their session, and it flags such broken sessions, stalled approvals, and context-window usage. The tool reads only local logs, requiring no proxy or daemon, and is under active development.

A Claude Code turn shows you its output, not its work. You send a prompt and the terminal shows a spinner. Behind it the window fills, subagents spawn and spend it on models of their own, the same context is read again on every call, and a compaction deflates the lot. Minutes later an answer arrives, and what it took to get there is not on screen. seedeep records the whole chain while it is still running — your prompt, every call to the model with its latency and its own input and output, every tool it fired and what came back, each subagent's work folded under the spawn that launched it, down to the answer you read. One continuous flow, live during the turn and still there to walk through afterwards. All of it from the logs Claude Code already writes on your machine: read-only, no proxy, no daemon, no session content sent anywhere. Your last turn looked like a spinner. This is what it was doing: 3% → 26% of the window, six subagents on three different models, 2.9M tokens billed — 2.5M of them the same context read again. Every line below is something your session does not tell you, and what seedeep does about it instead. A session died and never said so. A failed API call — an expired login, a session limit, an overloaded server — ends the turn and leaves it sitting there: measured over 1830 real transcripts, 39 of 47 failed calls were the last thing their session ever wrote. seedeep turns the tab red, files the session under Broken , and the menu-bar icon goes red above every other signal. A session has been waiting on you for ten minutes. An approval dialog reaches no log at all, so a stopped session looks exactly like a thinking one. seedeep reads Claude Code's own live state and turns the tab amber the moment it stops, saying what it is waiting to approve. A subagent is spending the window you needed. The live tree shows each one as it launches — its own context filling, the model it actually runs on, and the verbatim output it handed back to the main session. The window is not the size you assume. The bar follows the model your calls really run on, so /model mid-session moves it, and a Haiku subagent is measured against 200k while the session around it runs on 1M. The turn that paid to come back. A turn resumed after the cache went cold re-creates its whole prompt before doing any work — on a real corpus that is a quarter of every token spent , and it looks exactly like work. Waste you would otherwise reconstruct by hand. Seven deterministic checks no LLM score every turn as it closes, each quoting the Claude Code documentation that justifies it — and they name what the turn did right too. What the call actually cost. Every API call in the feed with its latency, its input and output on demand, and the split between cached context and what was new. What the session shipped. The commits and the tracker cards it produced, read from the calls that made them, never from a key typed in a prompt. seedeep is under active development. The complete tour of every surface → /duqaXxX/seedeep/blob/main/docs/features.md Every capture below is a synthetic session on a fictional project — no real path, prompt or project name appears in any frame. Each turn as a row: how many steps it took, how long it ran, whether a step failed, and the subagent rounds it spawned. It fills in as the session works — this is not a report produced afterwards. rules /duqaXxX/seedeep/blob/main/docs/trace.md A native menu-bar client, polling the same local server. It says what the session is doing right now, which subagents are running and on which model, and how full the window is — without a browser tab open. rules /duqaXxX/seedeep/blob/main/docs/tray.md Three things about a session are worth interrupting you for, and each has its own switch. Nothing else notifies: not a subagent finishing, not a tool error. A tray that notifies about everything is a tray you mute. | Banner | Ships | Why | |---|---|---| Waiting for your approval | on | the session cannot go on until you answer | The last API call failed | on | it has stopped, and nothing on screen says so — 39 of 47 real failures were the last line their session ever wrote | Turn finished | off | news you can read whenever you like should not be able to force you to silence the two that cannot wait | Each is one title and one line — which session, and what happened. The command awaiting approval and the error text stay in the panel: you cannot act on a banner, and the webhook channel sends its payload off your machine. seedeep ships unsigned, so macOS asks for the notification permission again on every update — a tray that goes quiet after an upgrade is that, not a bug /duqaXxX/seedeep/blob/main/docs/install.md installing-the-tray . Across every session on the machine: turn-size distribution, where the waste came from, and tokens split by the model that spent them — subagents counted under their own model , so a Haiku explorer inside an Opus session shows up as Haiku. Every word narrows. Your prompts and Claude's answers, matches highlighted in place, ranked by density rather than recency. Paste a commit hash or a tracker id and it also asks git and its own index — the session that did the work is exactly the one text search misses. rules /duqaXxX/seedeep/blob/main/docs/search.md Claude Code appends one line to a local session file per content block — a single response becomes several — each stamped with its call's token usage. seedeep tails those files and reconstructs the picture — no network interception, no ANTHROPIC BASE URL hack, nothing written back. It watches every active Claude Code session at once, and identifies its own launching session so it never counts itself. Session data flows one way: the server pushes to the browser over Server-Sent Events, and nothing seedeep reads is ever written. See docs/architecture.md /duqaXxX/seedeep/blob/main/docs/architecture.md for the full design. With Node or Bun, from npm: npm i -g seedeep or: bun install -g seedeep --trust seedeep watch, serve, and open the browser With neither , take the file for your platform from the latest release /duqaXxX/seedeep/releases/latest — macOS arm64/x64, Linux x64/arm64, Windows x64/arm64 — and run it. That file is not an installer, it is the program : it carries its own runtime and the whole browser GUI inside, installs nothing, and leaves behind only ~/.seedeep/ . The Linux builds require glibc Debian, Ubuntu, Fedora and derivatives — Alpine and other musl-based distributions are not supported. The menu-bar tray is a separate, optional download from the same release: a universal .dmg for macOS, a -setup.exe for Windows. It is a pure client — the server is what has to run where Claude Code runs. Both are unsigned , so macOS and Windows each show a first-launch warning. Inside Claude Code, seedeep install-command adds a /seedeep command that opens the GUI, stops the server, or reports what the current session cost. Installing, running, updating, remote access and removal in full → /duqaXxX/seedeep/blob/main/docs/install.md Everything above was checked by hand on macOS , on the machine seedeep is developed on. Linux has been used once , in a VM, on arm64; Windows in a VM , several times. Building for three systems is not the same as having used three, and this section is that difference written down rather than left for you to find. There is a middle ground worth naming: started is not used . Every release runs each server binary on a runner of its own operating system before anything is published — it must report its version, answer on its API and serve the browser GUI, or the release stays a draft. That rules out the download that dies at startup. It says nothing about whether the thing is pleasant, or correct, in front of a person on that machine. | macOS | Windows | Linux | | |---|---|---|---| Server — download or npm | Used daily; every claim above was checked here | Used on Windows 11 in a VM — installed from npm, server started and served its API against a real Claude Code session, status , stop , restart and install-command confirmed, and consecutive cold starts measured without a failure. /seedeep there needs one line of configuration install.md | arm64: used on Ubuntu 24 in a VM — GUI opened against a real Claude Code session, lifecycle and install-command confirmed. x64: exercised on every release, never used by a person — started, left idle and driven through the full lifecycle in CI, and version-checked on Docker; never run in front of a person. Both builds require glibc; Alpine/musl is not supported. | Tray | Used daily on a real menu bar | Installed and used on Windows 11 in a VM — the installer runs, the icon reads in the notification area, the popover opens at full height, trust-on-first-use and the connection screen work, the panel's buttons respond, notifications are delivered, and no console window appears | Not a target, deliberately: Tauri emits no tray click event on Linux, so the panel could not open — see docs/tray.md | Concretely: on Linux x64 you are the first to use it. A defect there is expected rather than surprising, and an issue /duqaXxX/seedeep/issues saying what you saw is the most useful thing you can send — every Windows session so far found several, which is the honest estimate of what an untouched square still holds. And which Claude Code hosts. Terminal sessions and the desktop app's Code tab are both watched live. One signal is missing on the desktop app: a session stopped at a tool approval reads as working rather than amber, because only a terminal session publishes that state and a transcript cannot tell a call awaiting your yes from one that is running. A question the model asks you does light amber there what each surface shows /duqaXxX/seedeep/blob/main/docs/features.md when-a-session-is-waiting-for-you . Read-only. seedeep only reads what Claude Code already writes. It never modifies, proxies, or intercepts your session. Live, not post-hoc. The point is watching a turn as it happens , not analyzing spend after the fact. Runtime-agnostic core. The reducer and every rule it applies are built on standard APIs alone — no runtime builtins — so they run and are tested anywhere. The server around them is Bun, and ships with it embedded: you never install a runtime to run seedeep . Local by default. Your session content never leaves the machine unless you ask for it, and asking for it turns on TLS and a token in the same move. The only outbound request seedeep makes on its own is the update check against registry.npmjs.org , and seedeep update --offline skips it. Visual is the point. The number is the message; the picture is what makes it click. The server and the browser GUI are developed against Bun https://bun.sh — no other runtime required. The menu-bar tray is a Tauri app, so building it additionally needs a Rust toolchain and the platform SDK. bun install bun start watch, serve, and open the browser bun run test run the test suite bun run typecheck tsc --noEmit the tests do not type-check bun run tray:dev build the tray panel, compile, run If you also run an installed seedeep, develop through bun run dev and bun run tray:dev : they give the checkout a state directory and a port of its own. CONTRIBUTING.md /duqaXxX/seedeep/blob/main/CONTRIBUTING.md has the full setup, the conventions, and how to send a change. features.md | install.md architecture.md api.md configuration.md · /duqaXxX/seedeep/blob/main/docs/trace.md trace.md · /duqaXxX/seedeep/blob/main/docs/search.md search.md tray.md session-output.md claude-code-upgrades.md CHANGELOG.md CONTRIBUTING.md SECURITY.md CODE OF CONDUCT.md MIT /duqaXxX/seedeep/blob/main/LICENSE © duqaXxX