Orrery

A local-first cockpit for your AI coding agents.

Every Claude Code, Kimi Code, DSH (DeepSeek), Codex, OpenCode, Z Code and Antigravity session on your machine in one window: tokens, disk usage, projects, subagents. Delete what you no longer need (Z Code is read-only), and route every harness through one local model proxy. Nothing leaves your computer.

English · 简体中文 · 日本語

[!NOTE]
The UI speaks English, 简体中文 and 日本語. It follows your system language, and you can switch anytime from the top bar.

Why

If you run several agent harnesses side by side, you hit the same friction every day:

1. Sessions are scattered. Each tool keeps its own JSONL or session folder, in its own format. You can't search, compare or resume across them.
2. Usage is invisible. How many tokens did that session really burn? How much disk are hundreds of transcripts eating?
3. Model switching is manual. Every harness has its own env vars and config files.

Orrery reads what the harnesses already write to disk and puts it all on one board. It doesn't wrap, fork or re-implement any agent.

What works today

How token counting works

Every harness records usage per API call, and none of them can simply be added up:

The session total adds up the main agent and every subagent. How it was checked:

• Claude Code: matches its own cost-state ledger exactly (input, cache read, output) over the same process window.
• DSH: matches DSH's own projection cache exactly up to the event the cache was built from.
• Codex: token_count and token_usage_record agree exactly in 8 of 12 files that have both; in the other 4 the difference is exactly the context-compaction calls.

Old Codex alpha sessions only stored a total without a breakdown. Orrery shows that part as unsplit instead of guessing.

OpenCode was checked per session against independent SQL and opencode stats: 53 records become 41 sessions plus 12 folded subagents. Input 108.4M, cache read 1853.0M and cache write 1.2M agree; output 4.7M includes 1.3M reasoning tokens. Session size measures UTF-8 payload bytes in message/part/event, not reclaimable SQLite file space. Windows desktop verified; macOS/Linux remain untested on hardware. Existing screenshots predate this adapter.

Known limits: the ledger resets on --resume, so Orrery rebuilds totals from the transcript instead. The ledger also counts side calls that never reach the transcript, such as title generation, so Orrery's Claude Code totals can read about 1–5% lower.

Getting started

Just want to run it? Grab the installer for your platform from the latest release:

Nothing is code-signed, so the first launch needs one confirmation: Windows SmartScreen → More info → Run anyway; macOS → right-click the app → Open, or xattr -cr /Applications/Orrery.app; Linux → chmod +x the AppImage.

[!NOTE]
Windows is what the author uses daily. The macOS and Linux builds are produced and tested by CI (clippy + unit tests on all three platforms), but nobody has run the app on a Mac or a Linux desktop yet. If something is broken there, an issue with the output of orrery from a terminal is very welcome.

Building from source — requirements

• Rust — MSVC toolchain on Windows
• Node.js 18+
• Windows: Visual Studio Build Tools with MSVC and a Windows SDK; WebView2 (preinstalled on Windows 10/11)
• macOS: Xcode command line tools
• Linux: libwebkit2gtk-4.1-dev libappindicator3-dev librsvg2-dev patchelf libxdo-dev

git clone https://github.com/arvelvale/orrery.git
cd orrery
npm install
npm run dev        # desktop app, reads your real sessions

Just want to look at the UI? No Rust needed:

npm run preview    # http://127.0.0.1:1420 with mock data

[!TIP]
Run Rust builds from PowerShell, not Git Bash. Git ships its own link.exe, which is not the MSVC linker.

Project layout

orrery/
├─ ui/                      # index.html · styles.css · app.js · i18n.js (shared by WebView and browser)
├─ src-tauri/
│  └─ src/
│     ├─ adapters/
│     │  ├─ mod.rs          # SessionSummary, TokenUsage, cache, shared helpers
│     │  ├─ claude_code.rs
│     │  ├─ kimi_code.rs
│     │  ├─ dsh.rs
│     │  ├─ codex.rs
│     │  └─ cleanup.rs      # session deletion and index cleanup
│     ├─ proxy/             # local model proxy
│     │  ├─ mod.rs         # start / stop / status
│     │  ├─ config.rs      # providers and routes (~/.orrery/proxy.json)
│     │  ├─ server.rs      # HTTP surface and upstream forwarding
│     │  └─ state.rs       # counters, last request, last error
│     └─ lib.rs             # Tauri commands
├─ scripts/preview.mjs      # zero-dependency static preview
├─ docs/                    # design docs (Chinese)
└─ DESIGN.md                # visual spec

Local model proxy

Point a harness at http://127.0.0.1:8787/v1 and Orrery forwards its requests upstream, so you can swap the model from the app instead of editing each tool's config.

# 1. put the key in the environment the app can see
setx ANTHROPIC_API_KEY sk-...        # Windows, then restart Orrery

# 2. start the proxy from the Models page, then point a harness at it
set ANTHROPIC_BASE_URL=http://127.0.0.1:8787

Providers, routes and the listen address live in ~/.orrery/proxy.json:

{
  "listen": "127.0.0.1:8787",
  "auto_start": false,
  "routes": { "cc": "claude-sonnet-4.6", "kimi": "kimi-k3" },
  "providers": {
    "anthropic": {
      "base_url": "https://api.anthropic.com/v1",
      "api_key_env": "ANTHROPIC_API_KEY",
      "wire": "anthropic",
      "model_prefixes": ["claude"]
    }
  }
}

Deleting sessions

Every session is just files on disk, so Orrery can remove the ones you no longer need. Nothing is deleted without a confirmation dialog that lists each session and the space it frees.

• Recycle Bin by default. Permanent deletion is a separate mode that needs an extra checkbox.
• Sessions in use are protected. Anything written in the last 10 minutes, or open in a running Claude Code, is skipped.
• Each tool's own index is cleaned too, so no dead entries are left behind. Index files are backed up to ~/.orrery/backups/ before they change.
• Codex goes through the official codex delete, which also clears Codex's history database. Orrery never writes to another tool's database.
• OpenCode goes through the official opencode session delete, subagents included. OpenCode has no Recycle Bin, so in that mode each session is first exported to ~/.orrery/exports/ with a RESTORE.txt that lists the opencode import commands. OpenCode's database file doesn't shrink right away; it reuses the freed space.
• Antigravity conversations open in a running agy are skipped. Only the conversation's own files are moved; agy's summary database is left alone, so its history list may keep the title.
• Paths are resolved by the backend from the session id and must stay inside that tool's data folder.

[!TIP]
Close the tool before deleting its sessions. A running Kimi Code or Codex may write the removed entries back into its index.

Privacy

• Orrery only writes to harness folders when you delete sessions, as described above.
• Its own files live in ~/.orrery/: your route config (including any API key you typed into the app, in plain text), index backups, and index.json — the parsed session list (titles, paths, token counts) that makes restarts instant. Delete it anytime; it is rebuilt on the next scan.
• No network calls and no telemetry. The proxy listens on 127.0.0.1 only.
• Fields that may contain pasted secrets (for example Kimi's lastPrompt) are never read.

Roadmap

• Tauri 2 shell and session hub
• Claude Code, Kimi Code, DSH and Codex adapters
• Token and disk accounting
• Local model proxy with real forwarding
• Persistent index for instant cold start
• Resume a session in its own CLI
• Tray, global shortcut, installer

Design docs, in Chinese: positioning · architecture · roadmap