Keep Claude Code prompt cache warm to cut API costs

claude-thermos

Stop paying to rebuild your Claude Code cache. When your main agent waits on a subagent for more than 5 minutes, its prompt cache silently expires, and the next turn re-encodes your entire conversation at the write rate instead of reading it back cheap. On long sessions with many subagents that's roughly 20% of your bill. claude-thermos keeps the cache warm so you never pay that tax.

Use

Run Claude Code exactly as you normally would, but through claude-thermos with uvx:

uvx claude-thermos                     # instead of: claude
uvx claude-thermos -p "fix the bug"    # any claude args pass straight through

Requires Python 3.11+ and the claude CLI on your PATH.

That's it. Warming runs automatically in the background. To disable it for a run without changing the command, set CLAUDE_THERMOS_DISABLE=1.

Tuning (all optional):

Choosing which claude to run

By default claude-thermos launches the claude found on your PATH. Point
it at a different binary with --bin, or the CLAUDE_THERMOS_BIN environment
variable. A bare name is looked up on PATH (so --bin claude-nightly
works); a full path is used as-is. This is handy for a vendored build or a
wrapper that exports a different CLAUDE_CONFIG_DIR per account.

claude-thermos --bin /path/to/bin/claude -p "fix the bug"
# or
export CLAUDE_THERMOS_BIN=/path/to/bin/claude
claude-thermos -p "fix the bug"

The flag must come before any passthrough claude args.

Daemon mode (shared proxy for the IDE and multiple terminals)

The default command warms only the claude process it launches. Clients that
launch claude themselves - the VSCode/Claude Code extension, which spawns
its own bundled binary - never go through it, and neither do other terminals.

claude-thermos serve runs the warming proxy as a standalone daemon on a
fixed loopback port. Point any client at it and they all share one warmer:

claude-thermos serve --port 8787          # run the daemon (Ctrl-C / SIGTERM to stop)

# then, for any client:
export ANTHROPIC_BASE_URL=http://127.0.0.1:8787
claude -p "fix the bug"                    # terminal — warmed by the daemon

For the VSCode extension, make sure its process inherits that environment
variable (on macOS, launchctl setenv ANTHROPIC_BASE_URL http://127.0.0.1:8787
before launching the app; or export it in the shell you start the editor from).
The extension honors ANTHROPIC_BASE_URL, so its traffic then flows through the
daemon and its main agent stays warm while subagents run.

The daemon observes traffic exactly like the launcher and already tracks many
sessions at once, so a single daemon serves every client on the machine. It
evicts sessions idle longer than --session-ttl (default 3600s) so it can run
indefinitely.

Tuning: serve accepts the same --idle/--interval/--max-cycles/--subagent-window
flags as the default command, plus:

Caveat: --upstream must be the real API, never the daemon's own loopback
address - otherwise the proxy would forward to itself. serve rejects a
loopback upstream, so if you export ANTHROPIC_BASE_URL globally, still start
the daemon with an explicit --upstream https://api.anthropic.com.

Why your cache keeps expiring

Claude Code's prompt cache uses a 5-minute TTL. Every turn, your whole conversation history is served from cache at 0.1x the input price instead of being re-sent at full price, as long as the cache stays alive.

The cache expires if more than 5 minutes pass between requests on the same prefix. The dominant trigger for that gap is not you thinking. It's the main agent blocked on a subagent that runs longer than 5 minutes. A subagent has a different system prompt and tool set, so its requests have a different cache prefix and never refresh the main agent's. While the subagent works, the main agent's cached history ages untouched; past 5 minutes it's gone. When the subagent returns, the main agent resumes with a byte-identical, append-only history, and finds its cache missing, forcing a full re-encode at the 1.25x write rate.

By then the history is large, so the re-encode is expensive: individual collapses re-write 200K to 500K tokens. Measured across roughly 185 local sessions, these rebuilds accounted for about 22% of the total bill, money spent re-encoding content that was already cached moments earlier.

How it works

claude-thermos launches Claude Code behind a small local reverse proxy (it points ANTHROPIC_BASE_URL at a loopback port; all traffic still goes to the real Anthropic API).

1. Observe. The proxy watches /v1/messages traffic and groups it into sessions and lineages, a lineage being one cache prefix, keyed by model + tool set + system text. The first tool-bearing lineage is the main agent; the rest are subagents.
2. Detect the danger window. When the main lineage goes idle and a subagent is actively running, the main prefix is at risk of expiring.
3. Warm. On an interval under the 5-minute TTL, it replays the main agent's last real request as a warm request: identical cacheable prefix, but max_tokens: 1 and no streaming. The single token is thrown away; the point is the prefill, which reads and refreshes the full cached prefix. Warm requests go directly to the API, never through the proxy, so they can't disturb real traffic.
4. Result. When the subagent finishes, the main agent's cache is still warm. It pays a cheap read instead of a full rewrite.

Each warm costs a cache read (0.1x); each rewrite it prevents would have cost a write (1.25x) on a much larger prefix, so the trade is heavily in your favor.

Event logs & savings

Every session writes to:

~/.claude-thermos/logs/<session_id>/
├── events.jsonl    # append-only structured event stream
└── summary.json    # rollup totals, written when the session ends

events.jsonl records each request/response's token usage plus every warming decision (warm_fired, warm_result, cap_reached, resume_detected, and so on). summary.json is the rollup you'll usually read:

All three cost figures are in base-input-token units (token counts already weighted by their cache multiplier). To turn net_savings into dollars, multiply it by your model's price per input token:

dollars saved ≈ net_savings × (input token price)

For example, at an input price of $3 / 1M tokens, a net_savings of 1_200_000 is about 1_200_000 × $3 / 1_000_000 = $3.60 saved that session.