Configuration

Configuration

Two files, both TOML, both safe to hand-edit while clauth runs (it reloads on external change):

• ~/.clauth/profiles.toml for everything program-wide: profile order, the active marker, the fallback chain, appearance, the scheduler.
• ~/.clauth/profiles/<name>/config.toml for one account: endpoint, key, env, model routing, its chain settings.

Every key below has a TUI equivalent on the Setup, Fallback, or Config tab (Interface and keys).

Account types

Claude Pro / Max / Team / Enterprise. Leave base_url blank. clauth captures the OAuth token from your session or mints one through clauth login, then detects the plan tier from Anthropic's profile endpoint.

API endpoint. Set base_url, and api_key if the endpoint wants one. Works against the Anthropic API or any compatible proxy. The key is handed to Claude Code through apiKeyHelper rather than written into settings.json.

Long-lived setup token. clauth login <name> --setup-token stores a claude setup-token mint as session-token.json. Sessions run on that static login, which never races clauth's token refresher. The Setup tab then shows a token row counting down to the re-mint.

The token outranks the profile's OAuth pair at every switch for as long as it exists, so a later clauth login <name> updates only what clauth polls usage with. clauth static-token <name> --clear drops it and puts the OAuth login back in front of sessions.

A mint is a narrower credential than a /login session: it carries user:inference and user:sessions:claude_code and no refresh token, against the five scopes a browser login stores. Claude Code turns off anything gated on the wider set, Claude in Chrome by name. Clear the token if you want those features back.

Third-party usage data

Three providers get typed usage panels:

Provider	Base URL	Shows
DeepSeek	https://api.deepseek.com	balance rows per currency: total, granted, topped up
Z.ai	https://api.z.ai	percentage bars per limit window (5h / 7d / 30d), per-tool rows, plan level, 7-day per-model token totals
Alibaba Model Studio	the four Qwen preset endpoints below	a 7d bar carrying your tier's absolute allowance, a 5h bar when the API reports one, plan tier, subscription status and days left

Any other endpoint is scanned best-effort: clauth probes a short list of usage paths on the origin your key already authorizes, and renders whatever percentage or balance shapes come back. Those panels carry a "looks wrong? report it" line, since the shape is guessed. An endpoint that returns nothing usable stops being polled until you press r.

The Alibaba console session

Alibaba is the one provider whose api key cannot read its own quota: every quota endpoint ignores the key outright. Those panels run on a separate console session instead. clauth login <account> on a Model Studio account opens the Alibaba console in your browser and stores the session it hands back as a [console] table in that account's config.toml. Nothing else on the account changes. The console returns an api key and an endpoint alongside the session, both scoped to a workspace rather than to your plan and billed separately from it, so clauth discards them.

The 48-hour clock runs from your aliyun console sign-in, so it is already ticking by the time clauth login captures the session. Running the login again inherits whatever is left of that window, which can be minutes; sign in to the Alibaba console afresh first if you want a full one. Once it lapses, clauth stops polling that account: the Usage tab reads console login expired, run clauth login and clauth list marks it (login expired). Polling resumes the moment a new session lands on disk, with no restart.

An account with no api key at all still gets its usage panel, since the quota rides the console session. It just cannot run Claude Code until you add a key.

Model routing

Per account, on the Setup tab or in config.toml:

[models]
default  = "opusplan"                     # preset alias or a full model id
opus     = "claude-opus-4-5-20251101"     # ANTHROPIC_DEFAULT_OPUS_MODEL
sonnet   = "claude-sonnet-4-5-20250929"   # ANTHROPIC_DEFAULT_SONNET_MODEL
haiku    = "claude-haiku-4-5-20251001"    # ANTHROPIC_DEFAULT_HAIKU_MODEL
fable    = "claude-fable-5"               # ANTHROPIC_DEFAULT_FABLE_MODEL
subagent = "claude-sonnet-4-5-20250929"   # CLAUDE_CODE_SUBAGENT_MODEL

default lands as the top-level model key in settings.json; the rest ride in its env block. A switch or a clauth start applies whichever account you land on.

Presets

A preset is a named base_url + [models] pair you can stamp onto any account from the Setup tab's a menu. Six ship built in:

Preset	Endpoint
DeepSeek	https://api.deepseek.com/anthropic
Z.ai	https://api.z.ai/api/anthropic
Qwen-TokenPlan-Intl	https://token-plan.ap-southeast-1.maas.aliyuncs.com/apps/anthropic
Qwen-TokenPlan-CN	https://token-plan.cn-beijing.maas.aliyuncs.com/apps/anthropic
Qwen-CodingPlan-Intl	https://coding-intl.dashscope.aliyuncs.com/apps/anthropic
Qwen-CodingPlan-CN	https://coding.dashscope.aliyuncs.com/apps/anthropic

DeepSeek and Z.ai set the endpoint plus a base model, leaving the tier rows yours to pin afterwards. The four Alibaba ones fill every row instead, because those endpoints reject a Claude model id outright rather than serving something for it, so any alias left unpinned fails on use. All six leave the api key alone; pick the region your plan was bought in, since a key issued for one is not accepted by the other.

save as preset stores the focused account's own endpoint and models under a name you type, in ~/.clauth/presets/<name>.json:

{ "base_url": "https://api.example/anthropic", "models": { "default": "my-model" } }

apply preset opens the picker, built-ins first. Applying replaces the account's endpoint and its whole [models] block, so a tier the preset leaves unset is cleared rather than kept; the picker warns and names the fields first when the account already carries any. The account's own api key is never touched, and a preset never carries one. d in the picker deletes a saved preset; the built-ins have no file and stay.

Auto-start the 5-hour window

The 5h window opens on a real inference call. clauth's own token refresh does not trip it, so an account can read 0% while the clock has yet to start.

auto_start = true    # per profile; older spelling kick_timer still reads

clauth then sends a 1-token Haiku ping on launch and on each refresh tick while no window is running. On a cold start it fetches usage before the first ping, so it never fires over a window that might already be live. That costs a fraction of a cent and it is a real billed /v1/messages call under your own token. Default off, OAuth accounts only.

If the messages limiter is blocking Claude Code, a live 5h window will not clear it. clauth re-tests with the same ping on the poll cadence and can rotate the chain around an account whose ping keeps getting rejected.

profiles.toml

Key	Type	Default	Controls
active_profile	string	none	the account currently linked into ~/.claude
profiles	list	[]	display order
fallback_chain	list	[]	ordered chain members (Auto-switch)
refresh_interval_ms	int	90000	usage poll cadence, 10 s to 1 h
refresh_spent_accounts	bool	true	keep polling accounts at 100%
preemptive_rotation	bool	true	rotate OAuth ahead of expiry; false waits for a rejection
weekly_switch_threshold	float	98.0	chain-wide 7d exhaustion line, 50-100
burn_aware_switching	bool	false	project usage forward instead of comparing to the threshold
burn_switch_floor_pct	float	98.0	earliest point burn-aware may switch, 90-100
burn_horizon_cap_ms	int	60000	how far ahead burn-aware projects
wrap_off	bool	false	switch off all accounts once the chain is out of quota
spend_budget_switching	bool	false	master switch for pay-as-you-go fallback
switch_off_when_budget_spent	bool	true	switch off once the spend ceiling is used up
auto_rescue	bool	false	lift an isolated run's transcripts into the global store before teardown
default_divergence	string	none	auto-resolve a credential mismatch: Overwrite, NewProfile, Discard
theme	string	auto	full or compatible
reset_display	string	relative	relative, clock, both
clock_format	string	24h	24h or 12h
show_estimates	bool	true	burn estimates on the Usage tab
show_pace	bool	false	ideal-pace marker on usage bars
count_cache	bool	false	count cache tokens in the Tokens totals
auth_broken	list	[]	accounts quarantined after a permanent OAuth rejection; clauth writes this

config.toml

Key	Type	Default	Controls
base_url	string	none	API endpoint; unset means an OAuth account
api_key	string	none	key for that endpoint
auto_start	bool	false	the 1-token window-opening ping (alias kick_timer)
disabled	bool	false	hide from auto-switch, polling, and the status feed
fallback_threshold	float	95.0	5h utilization % that switches away from this account
weekly_threshold	float	chain value	per-account override of the 7d line, 50-100
check_weekly	bool	true	count the aggregate weekly window against this account
check_scoped	bool	true	count per-model weekly windows against this account
last_resort	bool	false	the chain's parking spot
preferred	bool	false	the home account clauth returns to once it is clear
max_auto_spend	float	0.0	dollar ceiling on pay-as-you-go fallback
bell_threshold	float	none	5h % that fires a bell toast
[env]	table	{}	extra environment variables merged into settings.json while active
[models]	table	{}	default, opus, sonnet, haiku, fable, subagent
[console]	table	{}	Alibaba Model Studio usage session: token, site (international / domestic), region (ap-southeast-1 / cn-beijing). clauth login writes it (above)

last_resort and preferred are radio toggles across the chain: marking one clears it everywhere else, and no account can be both.

Storage layout

~/.clauth/
  profiles.toml            # everything in the table above
  price_cache.json         # LiteLLM model prices for the cost lens
  status_cache.json        # status.claude.com incident feed
  status.json              # the daemon's published snapshot (see Daemon)
  clauth.log, daemon.log   # event lines from the TUI and the daemon
  completions/             # generated shell completion scripts
  jobs/<id>.json           # backgrounded delegate jobs, GC'd after an hour
  live_sessions/<sid>.json # one row per live `clauth start` session
  profiles/
    work/
      config.toml          # everything in the table above
      credentials.json     # OAuth snapshot (.pending while a rotation is mid-write)
      session-token.json   # long-lived setup-token login, when captured
      usage_cache.json     # last-known utilization and plan
      usage_history.jsonl  # 2 days of samples, feeding burn-aware switching
      third_party_cache.json
      third_party_auth.json# set while the usage login is expired; a hash, never the credential
      account_id.json      # which account this is, so a re-login can be told apart
      profile_fetched.json # when the plan tier was last read
      kick_block.json      # messages-limiter block state
      throughput_cache.json# observed delegate tokens/sec per model
      runtime-<sid>/       # one CLAUDE_CONFIG_DIR tree per live session
      runtime-isolated-<sid>/
      sessions-<sid>/      # that session's PID file, flock-held while it runs

Lock files (.lock, clauthd.lock, usage-fetch.lock) sit alongside. Everything clauth owns is 0600, every directory 0700, re-tightened on each launch.

Deleting any *_cache.json, usage_history.jsonl, third_party_auth.json, or status.json costs you history and nothing else. Deleting credentials.json or session-token.json signs that profile out.

The -<sid> suffix appears only where the OS grants symlinks. Windows without symlink privilege builds the runtime tree by copying ~/.claude/, so every session of one profile shares a single unsuffixed runtime/ instead of paying for a copy each.