headroom

Your coding agents check their Claude and Codex limits before they start a job.

headroomd reads the 5-hour and weekly limits your subscriptions report and answers go, wait or reroute, over the CLI or MCP.

npm install -g headroomd   # or: brew install apertur3/tap/headroom
headroom setup
headroom can claude-fable --owner me && echo go

Then install the Claude Code plugin (requires headroomd above):

claude plugin marketplace add Apertur3/headroom
claude plugin install headroom@headroom

macOS, Linux, Windows (experimental). No telemetry. Zero runtime dependencies.

Not a cost tracker: ccusage estimates spend from local logs; headroomd reads the limit the vendor enforces. They work side by side.

Unofficial; some readers use undocumented vendor endpoints with your own login. See Vendor endpoints and your accounts.

Not related to headroomlabs-ai/headroom, the context-compression project.

Install

Node 22.13 or newer. The npm package is headroomd; the Homebrew formula is apertur3/tap/headroom. Both install the headroom command.

npm install -g headroomd
# or, on macOS and Linux: brew install apertur3/tap/headroom
# finds your logins, runs the doctor, asks before each change
headroom setup
# one block per account: percent used, reset time and pace state per window
headroom

For Claude Code, install the plugin after installing headroomd with npm or Homebrew:

claude plugin marketplace add Apertur3/headroom
claude plugin install headroom@headroom

It bundles the Headroom skill, the MCP server registration and a quota line on every prompt. Using it alongside headroom hook install does not print the quota line twice. See Claude Code plugin for requirements and uninstall commands.

Already running Claude Code or another agent? Copy skills/headroom/SKILL.md into its skills directory and say "set up headroom".

Without the plugin, headroom setup offers to register the MCP server for your default Claude Code profile. To register it by hand, or in another profile:

claude mcp add --scope user headroom -- headroom mcp
CLAUDE_CONFIG_DIR=~/.claude2 claude mcp add --scope user headroom -- headroom mcp

Agents without MCP, Codex for example, call the CLI. Then give your agents the instructions in For agents. The MCP tools and example calls are in docs/mcp-and-agents.md.

The three commands above are the short path. Discovery on its own:

headroom accounts discover   # finds Claude, Codex, Antigravity, Grok and Kimi logins

On macOS and Linux, brew install apertur3/tap/headroom installs the same package and adds a brew services start headroom service.

By hand, the same steps are one command:

headroom setup                # discovery, doctor, service, MCP registration -- asks before each change

Run headroom notify configure to choose notification channels, a preset and overrides. Run headroom update only when you choose to install a newer package; headroom uninstall reverses setup, and headroom completion <bash|zsh|fish|pwsh> prints a shell completion script.

accounts discover prints what it wrote (Wrote ~/.headroom/accounts.toml (4 accounts). Next: headroom doctor) and, the first time, seeds ~/.headroom/policy.toml and routing.toml from examples/ so headroom can <class> --owner <your-agent-name> works immediately with the example action classes (claude-fable, codex-build, gemini-bulk) -- edit routing.toml to match your accounts.

Keep an optional second profile that is deliberately logged out without polling it: add enabled = false to its [[accounts]] block, or run headroom accounts disable <name>. It remains configured and can be restored with headroom accounts enable <name>.

Full walkthrough, including what each step grants and why: docs/quickstart.md.

status/doctor check the npm registry at most once a day (a plain GET of the headroomd version, no account data) and print a one-line notice when a newer headroomd is out; run headroom update (never automatic) to install it. Set update_check = false in ~/.headroom/policy.toml to turn the check off -- see Staying up to date.

Related MCP server: mcp-token-saver

For agents

Paste this into the CLAUDE.md or AGENTS.md your agents read, and replace the placeholders with your own action classes, meters and agent name. The agent then checks Headroom before it dispatches work, instead of asking you.

## Usage limits (Headroom)

Check Headroom before you start work on a subscription. Do not ask me about usage or reset times.

- `headroom can <class> --owner <your-name>`: exit 0 means go, exit 2 means no. Any other exit
  is an error: treat it as no. Classes are the `[consumes]` entries in `~/.headroom/routing.toml`.
- On no, try the next class in your fallback list. If none has room, run
  `headroom wait --meter <meter> --until-reset --max 6h` (exit 3 means it timed out), then run
  `headroom can` again before you start anything.
- Start a job with
  `headroom run --meter <meter> --need <window>:<points> --owner <your-name> -- <command>`,
  for example `--need 5h:10` for 10 percent of the 5-hour window. It reserves the points, runs
  the command and releases them when it exits. Exit 2 means refused; otherwise it returns the
  command's own exit code.
- UNKNOWN means there is no trustworthy reading. Treat it as no. Never pass `--allow-unknown`
  unless I ask for it.
- For the full picture, read `headroom --json`, not the human view.
- In Claude Code with the Headroom MCP server, `quota_can` answers the same question as
  `headroom can`.

To put live quota in front of the agent on every turn, run headroom hook install --agent claude (or answer yes in headroom setup). Each Claude Code prompt then carries one line the daemon wrote at its last poll, with its age, for about 60 tokens per turn; headroom line prints the same line for any agent. Claude Code on macOS and Linux only for now; Windows, Codex and Gemini hooks are not supported yet.

A plain can is advisory. A reservation tells other agents that use Headroom the points are taken; it does not limit what the job itself spends, and it cannot stop tools that do not use Headroom. The longer version, with pace states, leases and reserves, is skills/headroom/SKILL.md.

Platform support and known limitations

Why

An agent that fans out several jobs needs two answers first: can this account afford them, and which other jobs have already reserved capacity? Vendor screens answer neither for a program, and a stale reading can send more work to an account that no longer has capacity. Headroom combines vendor readings with cooperative reservations. When a reading is stale or failed it says UNKNOWN instead of a number, and can answers NO for it unless you pass --allow-unknown.

One daemon reads quota across Claude, Codex, Antigravity, Grok, Kimi, and local vLLM or llama.cpp pools, keeps history, and detects resets. Antigravity and Grok support is experimental.

Demo

How it differs from a usage analyzer

Usage analyzers such as ccusage read local usage files and report estimated costs. Headroom reads the usage and limits your vendors report, from their own usage endpoints, and answers before work starts: can, gate, route, leases and an inbox, for agents that dispatch work. They answer different questions and work side by side.

Before and after

Without a check, an agent finds the limit by hitting it, in the middle of a task. With Headroom it checks first and tries the next class in your fallback list when the answer is NO. can only checks; when several jobs start at once, use headroom run so each one reserves its capacity.

Without Headroom

# weekly limit reached mid-task, work stops half done
codex exec "refactor the billing module"

With Headroom

if headroom can codex-build --owner builder; then
  codex exec "refactor the billing module"
elif headroom can claude-fable --owner builder; then
  claude -p "refactor the billing module"
else
  # no class has room: wait for a reset, then check again before dispatching
  headroom wait --meter codex-main:main --until-reset --max 6h
fi

headroom run --meter ... --need ... --owner ... -- <command> goes one step further: it checks, reserves the capacity for the job's duration so other agents that use Headroom see it as taken, runs the command, and releases the reservation. Runnable examples against synthetic data are in examples/.

Features

Headroom is not a router. Which model is good at what is your opinion and changes monthly. Keep it in ~/.headroom/routing.toml; Headroom only filters your fallback list by budget. It also never sits in the request path.

Dashboard

Synthetic data, generated by scripts/render-dashboard-demo.mjs.

Human dashboard and browser report

Open the terminal overview, or export a local browser snapshot:

headroom dashboard
headroom dashboard --once --ascii

Use Tab to select a meter, v for details, and e for events. Both chart views show recorded remaining capacity, with gaps where readings are missing or untrusted.

To open the same recorded history in a browser:

headroom dashboard --html report.html

Open the generated file in your browser. The report shows subscription usage and reset times in an overview, with selectable window charts and sample details. It supports light and dark themes and makes no external requests. It is a snapshot: generate it again for newer readings. Existing files are preserved unless you pass --force; new files use mode 0600 on systems that support Unix permissions.

Design principles

Each of these is a rule the code enforces and a test pins. Past real-world breakage is the reason for every one.

• Fail closed: never a fake number. A missing, stale or inconsistent reading becomes UNKNOWN, and can answers NO for it unless you explicitly pass --allow-unknown.

• A stale-lane canary that cannot be silenced. A lane with no fresh reading for 6 hours raises an alert, whatever the notification config says.

• Only kill what Headroom provably started. On macOS and Linux, a process is signalled only if its recorded pid, command and start time still match.

• Secrets stay in their own stores. Notification tokens live in the OS secret store; vendor credentials are read at use time from the vendor's own store or credential file and never copied. Output is redacted.

• A redacted fixture per past breakage. Every Antigravity failure shape seen in the field has a fixture and a test.

• Fail closed: never a fake number. A missing, stale, held or inconsistent reading becomes UNKNOWN, and can answers NO for it unless you pass --allow-unknown (freshnessGate, canConsume, tested in store-policy.test.ts). Antigravity lanes that report no usage are classified as blocked, idle or unknown, never filled in (antigravity-lanes.ts, antigravity-lane-classifier.test.ts).

• A stale-lane canary that cannot be silenced and does not spam humans. A meter lane with no fresh reading for 6 hours (configurable) raises an alert into the inbox and headroom doctor. It runs before and independently of the notification config, so a muted channel or events_off cannot hide it. External channels get nothing per lane; at most one plain-language message per principal per episode, after 24 hours, held during quiet hours (canary.ts, stale-lane-canary.test.ts).

• Process hygiene: only kill what Headroom provably started. A process is signalled only if recorded evidence (pid, command and start time from when Headroom launched it) still matches the live process, re-checked immediately before every signal, so a recycled pid is never hit. An age watchdog reaps Headroom-started agy processes that outlive their limit (process-tree.ts, agy-watchdog.ts, antigravity-keepalive-stop-recycled-pid.test.ts, agy-watchdog.test.ts). The test suite itself fails if any process it started outlives the run (global-leak-gate.ts). The watchdog and the pid and start-time checks are not active on Windows.

• Secrets stay in their own stores. Headroom's own credentials (notification tokens) are read at use time from the macOS Keychain (secret-tool on Linux, Credential Manager on Windows) and never written to disk; vendor tokens are read at call time from the vendor's own store or credential file (for Kimi's manual token and the Moonshot key, a 0600 file you save yourself) and dropped. Output and logs are redacted (notify.ts, security.ts, security.test.ts, SECURITY.md).

• A redacted fixture per past breakage. Every Antigravity failure shape seen in the field has a fixture (recorded and redacted, or synthetic) and a test that runs it through the real code path (test/fixtures/antigravity/, record-antigravity-fixture.sh). The repo also scans itself and its packed npm tarball for private addresses, emails and home paths (privacy-sweep.sh, run in CI).

How it fits together

Headroom reads Claude, Codex, Grok and Kimi through native TypeScript adapters. Antigravity is experimental and supported on macOS 14 or later: it requires a logged-in agy and a running Headroom daemon. The npm and Homebrew packages include a universal macOS reader, verified by SHA-256 before use. Gemini CLI is not required or supported for consumer subscriptions; Google ended consumer Gemini CLI access on June 18, 2026.

For Antigravity quota, the bundled macOS reader finds the running Antigravity language server (the daemon-kept agy) in the process list, takes its local port and, where the process has one, the local CSRF token from its command line, and asks that server on 127.0.0.1 for its quota summary. It does not read agy's OAuth or session credential. The daemon also reads the tail of agy's newest log under ~/.gemini/antigravity-cli/log for login-state markers only. The optional Antigravity model-list check is off by default and is the one path that uses the Gemini CLI's OAuth file; see Vendor endpoints and your accounts. Missing or failed summaries stay UNKNOWN. An idle window with real fractions may carry a doubt marker when its reset time looks synthetic. A five-hour window entirely absent from an otherwise-successful response (a genuinely idle rolling window) prints n/a instead of UNKNOWN and does not block a budget check. The endpoint contracts and native reader build on CodexBar (MIT).

A budget check before a job

This synthetic example has 20% usage and another owner reserving 75%. The remaining capacity cannot cover another 10-point job, so the child command never starts:

$ headroom run --meter claude-main:all --need wk:10 --owner builder -- codex exec "Run the tests"
wk needs 10 more but only 0.0 left after 75.0% already leased before the 10% reserve

When there is enough capacity, run reserves it, starts the command, and releases the reservation when it finishes. The check and reservation are atomic across processes. A plain can or gate is advisory; use run or can --lease when starting work. These reservations coordinate agents that use Headroom; they cannot stop unrelated tools from consuming the subscription.

Examples

Runnable against synthetic data, no logins needed. Build once, then point the CLI at a seeded home:

npm run build
# sets HEADROOM_HOME to a throwaway demo home
eval "$(node scripts/demo-home.mjs)"
# walks a fallback list, dispatches on the first class with room
bash examples/orchestrator-loop.sh
# refuses a job the budget cannot cover (exit 1)
bash examples/ci-gate.sh
# the MCP call a client sends, over stdio
bash examples/mcp-quota-can.sh

The scripts refuse to run unless HEADROOM_HOME points at that demo home, so they cannot touch a real ~/.headroom, and they call the repo build (dist/cli.js), not a global headroom. They print which binary they use.

See examples/ for the scripts and the starter accounts.toml, policy.toml and routing.toml.

Documentation

• docs/claude-code-plugin.md: Claude Code plugin install, quota line, MCP and uninstall

• docs/quickstart.md: install to first truthful line, macOS, Linux and Windows

• docs/concepts.md: principal, meter, window, observation, pace states, leases, events

• docs/mcp-and-agents.md: the MCP tools, example calls, and how an orchestrator should use them

• docs/json-contract.md: the versioned field-by-field shape of every --json output and MCP tool result, and its compatibility promise

• docs/vendors.md: what Headroom reads per vendor, and its known live limitations

Vendor endpoints and your accounts

Several adapters read usage endpoints that the vendors do not document as public APIs. Headroom calls them with your own credentials, taken from the vendor CLIs you installed and signed in to on the same machine, and only to read usage: it sends no prompts, changes no account settings, and never writes to another tool's credential store. Use it only with accounts you own. Vendor terms may restrict this kind of access, and a vendor can change or block an endpoint at any time; use it at your own risk. Headroom does not rotate credentials to get past a block or a limit.

Per-vendor detail: docs/vendors.md.

Security

Headroom never writes a vendor credential or notification token to disk, and redacts secrets from its output. On macOS a bundled helper, headroom-claude-probe, reads Claude Code's Keychain token by running /usr/bin/security, which the Keychain item's own access list admits, keeps it in the helper's memory and makes the usage request itself, so the token never reaches Node or Headroom's output. With Claude Code 2.1.263 on macOS 26.5 this needed no Keychain dialog. The helper ships inside the npm package as a universal binary, verified against a recorded SHA-256 before every use. Tokens are otherwise read at call time from the vendor's own credential file and dropped after the request. The daemon listens on a 0600 local socket on macOS and Linux. On Windows it uses a named pipe where both sides prove they hold a per-daemon session token, kept in a 0600 file under HEADROOM_HOME. There is no telemetry. Apart from the vendors you configure and the notification channels you enable, the only request Headroom makes unprompted is the daily npm version check above, a plain GET with no account data (update_check = false turns it off); headroom update and headroom engine install contact npm and GitHub only when you run them. The engine is pinned and checksum verified, and every query lands in an audit log. Details in SECURITY.md.

Status

Releases are published to npm as headroomd; CHANGELOG.md lists what changed in each one. Vendor endpoints are private and change without notice: Headroom pins them, records fixtures, backs off on 401, 403 and 429, and prints UNKNOWN instead of a stale number. What is and is not verified is in Platform support and known limitations. Antigravity requires macOS and a daemon-kept agy; see vendor setup.

Contributions are welcome: see CONTRIBUTING.md and the issues labelled good first issue.

MIT. Third party notices in THIRD_PARTY_NOTICES.md.