UsageMax public observability

UsageMax

Your AI work, in perspective.

A private-by-default usage ledger for people and teams building with AI.

Live app · Docs · Integrations · Pricing · How we count · Issues

UsageMax reconciles the AI usage histories already on your computers into one clear view of tokens, model mix, tracked cost, sessions, and activity. The open-source CLI performs bounded local scans and uploads aggregate snapshots to the versioned UsageMax API.

The short version: one-shot collector, private by default, public only by

choice. Prompts and completions never leave the computer.

What you get

The physical HUD/screen project is intentionally separate from this repository. It can consume UsageMax telemetry, but it is not required to use the platform.

Connect a computer

1. Create a link code

Sign in at usagemax.com/account, choose Link a computer, give it a name, and copy the one-use command. The code is valid for ten minutes and can be used once.

2. Link and sync

Run the command on the computer or WSL distribution that owns the history:

bunx usagemax link UMX-XXXX-XXXX-XXXX-XXXX

Optional explicit name; the account keeps this name for the installation.

bunx usagemax link UMX-XXXX-XXXX-XXXX-XXXX --name "Work laptop"

The first link performs a full one-shot sync. Every installation receives a stable random ID, so relinking or renaming it rotates its key without creating a duplicate device.

3. Reconcile when you choose

bunx usagemax sync --dry-run --explain  # inspect the bounded plan

bunx usagemax sync # upload changed usage bunx usagemax sync --full # reconcile all retained history bunx usagemax status --remote # check the saved credential safely

CLI updates and progress

The CLI checks npm's latest dist-tag at most twice per day. Run usagemax update to update a global installation through Bun or npm; use --check for a read-only check. Run usagemax update sync to hand a command to the current release without typing @latest, or set USAGEMAX_AUTO_UPDATE=1 for an explicit automatic handoff. Use --no-update-check or set USAGEMAX_DISABLE_UPDATE_CHECK=1 in offline environments. Interactive terminals show a small stderr progress line; JSON, quiet, CI, and scheduled runs remain machine-readable and quiet.

Coverage and correctness

UsageMax pins ccusage v20.0.24 and uses its 16 adapters: Amp, Claude Code, Codebuff, Codex, GitHub Copilot CLI, Factory Droid, Gemini CLI, Goose, Grok Build, Hermes, Kilo Code, Kimi CLI, OpenClaw, OpenCode, Pi, and Qwen Code. Named Pi-format stores are discovered as well.

The collector recognizes supported provider overrides, bounded home locations, Claude Desktop sessions, .cc-mirror, renamed Claude/Codex backup folders, and supported Windows homes from WSL. Normal syncs inspect known locations and immediate home entries; they do not crawl the whole disk.

bunx usagemax doctor                 # metadata-only coverage check

bunx usagemax doctor --deep --json # retained-history audit bunx usagemax sync --archives # one-time compressed-history recovery bunx usagemax report # local ccusage report bunx usagemax status --json # version, coverage, checkpoints, and link state bunx usagemax update --json # machine-readable npm release check

Full scans can catalog retained history from 2024 onward. A readable file inventory is not proof that an aggregate parser consumed every file and day. UsageMax reports those two states separately; counter decreases and missing rows remain protected unless the parser explicitly certifies its coverage. Unknown models stay unattributed; unknown pricing stays unknown.

Cursor, Windsurf, Aider, Continue, Cline, Roo Code, hosted agents, direct provider API traffic, and enterprise billing systems do not all expose a stable local ledger. Feed those through the native or OTLP/HTTP JSON contract, or use a provider billing export when local evidence is unavailable.

Privacy and resource use

The CLI is short-lived. It skips unchanged inventories, parses only the necessary date range, uses a local lock to prevent overlap, and never downloads a package per run. Optional scheduling invokes the same one-shot process and backs off after failures.

bun install -g usagemax

usagemax service install # approximately every 15 minutes usagemax service install --every 30 usagemax service status usagemax service uninstall

Scheduling is opt-in. macOS uses a user LaunchAgent, Linux/WSL uses a user systemd timer, and Windows uses Task Scheduler. Sleeping or battery-powered computers are not needlessly woken.

Credentials and safe diagnostics

UsageMax has separate credentials for website sign-in, one-use linking, and collector writes. A collector key is exactly umx_ followed by 64 lowercase hexadecimal characters. It is generated server-side, displayed once, stored locally with user-only permissions, and stored by UsageMax only as a SHA-256 hash.

For an Advanced · custom telemetry collector key, pipe the secret through stdin. Never put it in a command-line argument, URL, request body, repository, or log:

set +x

set -o pipefail printf '%s' "$USAGEMAX_COLLECTOR_TOKEN" \ | bunx usagemax token status \ --json \ | jq -r '[.httpStatus, (if .ingestAuthorized then 1 else 0 end)] | @tsv'

token status and telemetry test are included in CLI 0.3.11. If a fresh environment still has an older npm tag, run usagemax update token status or use node packages/cli/src/cli.js token status from this repository until the new package is published.

To prove the authenticated write path, send one content-free agent_state event with all token and cost values set to zero:

set +x

set -o pipefail printf '%s' "$USAGEMAX_COLLECTOR_TOKEN" \ | bunx usagemax telemetry test \ --token-stdin \ --json

This writes one observability event and does not change token or spend totals. An unbound advanced key becomes bound to this CLI installation on its first accepted write. The CLI returns a non-zero exit code if the credential cannot ingest or the API rejects the event.

The numeric projection is deliberately small:

The read-only diagnostic reports credential type, scopes, activation state, profile/name, binding state, and last accepted/rejected write. It never returns the token, its hash, or the authorized UUID. Advanced keys are active immediately; they bind on the first valid write. Linked CLI keys are bound during the link exchange.

Agent and developer surfaces

Public reads are bounded and unauthenticated. The public MCP servers are stateless, read-only, and do not expose private workspace data or collector credentials. Collector ingestion is a separate authenticated API boundary.

Repository map

src/             Next.js web app, public routes, docs, and UI

convex/ Convex schema, authorization, ingestion, and projections packages/cli/ Distributable usagemax package docs/ Architecture, contracts, coverage, operations, and releases skills/ Direct-installable agent skills and source notes server.json MCP Registry submission metadata for the public remote server

For the documentation map, start with docs/README.md. It separates public product and integration guidance from contributor, operator, release, and historical records.

MCP Registry and skills.sh status

The public MCP server is published in the official Registry as io.github.SYMBaiEX/usagemax@1.0.0. The server.json file is the versioned source metadata for that listing; it contains no package or credential claim. The three `ski