Providers and cost

vir makes two model calls per session: a cheap classify (Haiku) and a distill (the main cost). Which provider handles them is one config key.

Providers

`provider`	Auth	Cost model	Notes
`anthropic` (default)	`anthropicApiKey`	Per-token, Anthropic list price	Predictable; no effect on Claude Code limits
`claude-cli`	none — uses the installed `claude` binary	$0; consumes your Claude Code subscription quota	Capped at 25 sessions per run; halts cleanly at a subscription limit with the reset time
`codex-cli` (experimental)	none — uses the installed `codex` binary and your ChatGPT login	$0; consumes your Codex usage limits	Same 25-session cap and halt-at-limit as `claude-cli`; Codex picks the model unless you pin one. See Codex
`kie`	`kieApiKey`	Per-token, ~28% of Anthropic	Third-party proxy. Kept for cheap testing and calibration runs, not recommended for your real vault; `kieTopUpTier: "high"` applies the 10% bonus-credit discount

Which one? anthropic for predictability, claude-cli if you already pay for Claude and don’t mind distills sharing its quota, codex-cli likewise for a ChatGPT plan. Kie exists because it’s cheap enough to burn through hundreds of sessions while calibrating prompts — that’s what it’s used for in vir’s own development. It routes your transcripts through a third party, so it isn’t the right choice for the vault you actually rely on.

claude-cli always runs with --no-session-persistence and a neutral working directory, so vir never writes transcripts into ~/.claude/projects (no self-scanning) and no project’s CLAUDE.md leaks into distill context. codex-cli does the same with --ephemeral. Both run with no tools: the distill prompt carries transcript text vir doesn’t control, so claude -p gets --tools "" --strict-mcp-config (no built-in tools, no MCP servers or connectors) and codex exec runs read-only with its shell, web search, apps, plugins, MCP servers and subagents switched off. Calls on both are logged with cost marked not-applicable — never a fake $0.00.

Measured cost

From the author’s machine, six months, Anthropic API:


Sessions distilled	297
Total	$25.85
Median per session	$0.004
90th percentile	$0.14

Long sessions with hundreds of tool calls are the tail. vir strips large tool outputs and oversized skill loads before distilling; on one 517-tool-call session that cut input from ~217k to ~95k tokens without losing signal.

Hybrid routing

Haiku captures routine and tool-heavy sessions as well as Sonnet at about a third of the cost; it only misses higher-order lessons on decision-heavy or very large sessions. When models.distillFast is set, each session routes after classification:

category === "decision" → models.distill
input tokens > models.distillThreshold (default 100,000) → models.distill
otherwise → models.distillFast

vir init enables this by default. Existing installs are untouched on upgrade: with distillFast unset, models.distill handles everything as before.

Cost controls

vir run --dry-run              # per-session estimate, exit before any call
vir run --force-model haiku    # override the distill model for this run
vir cost                       # actuals from ~/.vir/cost.log: total, median, p90
vir cost --since 30d --top 10
vir cost --by-session

vir run asks before proceeding with more than 20 new sessions (--yes skips). Every recorded cost is real token usage when the provider reports it, else a chars/4 estimate, and the record says which. Pricing is provider-aware and overridable under pricing in config.