agentglass — mission control for every AI coding agent on your machine

Workspace

Press a key. Fly the cockpit.

Each panel below is the real layout of the app — and the keys are real too. Try it: press d, g, o, t or c on your keyboard right now.

WORKSPACE Diff — everything the fleet changed one keystroke away

src/services/pricing.tsM · 2 chunks

41 const rate = taxRate(region)

42- const price = subtotal * rate

42+ const price = round(subtotal * rate)

43 cart.total = price // word-level diff, shiki-highlighted

17 export function checkout(cart: Cart) {

18+ assertRounded(cart.total)

19 return pay(cart) }

every Edit/Write the fleet made, chaptered per session · j/k move · v split/unified · check off reviewed files

shop-api · main↑2 ahead

Staged (2)

M src/services/pricing.ts

M src/routes/checkout.ts

Unstaged (1)

M web/src/components/Cart.tsx

Commit

fix: round prices at checkout

lazygit-style, write-gated · s stage · u unstage · x discard · interactive hunk staging · branches, log, stashes

compose: shop

api-workerrunningcpu 34%212 MB

postgresrunningcpu 12%148 MB

redisrunningcpu 3%18 MB

logs — api-worker

13:46:02 INFO worker started · queue=payments 13:46:04 INFO processed 14 jobs in 412ms 13:46:07 WARN retry #1 · stripe timeout (2.1s) 13:46:08 INFO ok · job 8842 settled $129.00

lazydocker-style, write-gated · r restart · l logs · s stop · images, volumes, networks

dev@machine shop-api % make test bun test v1.2.19 ✓ pricing rounds correctly (12ms) ✓ checkout asserts rounded totals (8ms) ✓ cart merges line items (21ms) 3 pass · 0 fail · 41 expect() calls dev@machine shop-api %

a real PTY — your login shell over a WebSocket, not an emulation · per-repo, persistent · ⚙ discovers Makefile targets & package.json scripts

shop-webOpusacceptEdits

fix the cart rounding bug

Read Cart.tsxGrep pricingEdit pricing.ts

Found it — totals were summed before rounding. Patched round() into pricing.ts and added a guard in checkout. Running the tests now.

drives your local claude CLI · pick repo, model, permission mode · sessions join the fleet like any other agent

Sessions

Departures — every agent session

Every session on your machine, live. History appears the moment you open the cockpit, and it persists across reloads.

LIVE Fleet board 18 live · 6 projects

Flight	Model	Origin	Tools	Tokens	Spend	Remark
AG-7A3F	Opus	shop-web	3	183.8k	$211.61	on time
AG-E2B8	GPT-4o	shop-web	5	1.36M	$160.21	running Bash
AG-3C9A	Sonnet	shop-api	12	930.9k	$469.83	on time
AG-D4E9	Gemini Flash	sandbox	6	18.0k	$9.70	running Read
AG-5F6D	Opus	payments-svc	9	1.25M	$571.22	✋ holding — approve?

18 sessions tracked 6 projects spend window $359.85 sources: claude code hooks · any otel genai exporter · transcript scanner

Telemetry

And the dashboard never blinks.

While you fly the panels, the glass keeps watching every provider — persisted across reloads.

Cost

Where the money goes

Cost per model, session and project. Tool-latency p50/p95, error timelines, session lifecycles.

Opus 352.0k $188.00 · GPT-4o 275.2k $94.42 · Sonnet $34.42

Alerts

What needs you

Dangerous tool calls hold for your clearance. Approve or deny remotely — opt-in, audited.

✋ Approve Bash? kubectl delete deploy api [Approve] [Deny]

Providers

Any aircraft, any fleet

Claude Code hooks, or any OpenTelemetry GenAI exporter: Codex, Gemini CLI, Bedrock, LangChain, LiteLLM.

otel/genai ▸ rx 1.33 events/sec · peak 2/s

Quickstart

Zero to cockpit in 90 seconds

Runs on localhost — your data never leaves your machine.

pilot ▸ request startup.

tower ▸ cleared. run npx agentglass

◇ server listening on :4177

◇ scanning ~/.claude/projects … 18 sessions found

◇ claude code hooks installed

pilot ▸ tower in sight.

tower ▸ a note on security —

◇ local-first: SQLite on your disk, nothing phones home

◇ the control plane is opt-in and audited

◇ MIT license · Bun server · React UI · Tauri desktop app

◇ read SECURITY.md before installing — we mean it

tower ▸ github issues welcome, frequency monitored.

EVERY AGENT