Small, sharp, local-first tools for people who ship agents and need to see what they're actually doing.
Sharp, local-first tools — lint your tools, capture a run, replay it, and fix what's slow.
| what it does | |
|---|---|
🧹 toollint  |
Linter for tool / function-calling schemas — catches the mistakes that make agent tool-calls silently fail. Anthropic + OpenAI dialects, zero deps. |
🚰 llmtap  |
Transparent proxy — record every Anthropic API call with zero code change (one env var). Streaming-aware, live cost/cache dashboard. |
⏪ rewind  |
Time-travel debugger — inspect any step of a recorded run, then fork & replay it with edits to ask "what if?" |
🔍 cachelens  |
Prompt-cache doctor — finds why your cache isn't hitting and how much it's costing you. |
💧 trickle  |
Best-effort parser for partial / streaming JSON — read tool-call fields as they stream in, before the closing brace. Zero deps. |
✂️ ctxfit  |
Trim conversation history to a token budget without breaking it — never orphans a tool_use/tool_result pair. Zero deps. |
✅ msglint  |
Pre-flight validator for the Anthropic Messages API — catches the structural bugs that cause 400s (bad alternation, dangling tool_use, system-in-messages) before you send. Zero deps. |
🕶️ cloak  |
Reversibly redact secrets & PII from prompts before they leave your machine — keys, emails, cards, IPs → stable placeholders, restored when the model echoes them back. Zero deps, offline. |
🌀 spincheck  |
Catch an agent stuck in a loop — repeats, A-B-A-B cycles, low-novelty churn — before it burns your token budget. Watches the tool-call stream, trips with a reason. Zero deps. |
🔧 coax  |
Repair & coerce tool-call arguments to match your JSON Schema — "5"→5, "yes"→true, stringified-JSON→object, scalar→array, wrong-case enums. Fixes the type wobble that throws in your handler, with an audit trail. Zero deps. |
🥫 canned  |
Deterministic, offline test double for the Anthropic Messages API — script replies (tool calls included), drop the stub in where your code wants a client, assert on what the agent did. No API key, no network. Zero deps. |
📼 streamfold  |
Reconstruct the final Anthropic message from a recorded SSE stream — folds text/tool_use/thinking deltas + usage back into one message, and flags truncated or corrupt streams. Offline, zero deps. |
🧾 tariff  |
Cost calculator for Anthropic API usage logs — turns logged usage into dollars with the cache tiers done right (read 0.1×, write 1.25×/2×, batch 0.5×). Digs records out of messy logs, groups by model. No key, offline, zero deps. |
🧬 schemize  |
Infer a JSON Schema from example JSON values — feed it a sample tool output or API response and get a clean draft-2020-12 schema for your tool params / structured outputs. Merges examples (required vs optional, null/number unions), detects formats. Offline, zero deps. |
🔁 transmute  |
Convert chat payloads between OpenAI and Anthropic — moves the system prompt, rewrites tool_calls↔tool_use, role:"tool"↔tool_result, images and tool defs, and fixes role alternation so the other provider just accepts it. Both directions, offline, zero deps. |
🔬 promptdiff  |
Structure-aware diff of two Anthropic Messages API requests — skips the JSON-formatting noise and shows what actually changed: model/params, tools added/removed/retyped, a word-level system-prompt diff, and message inserts/edits. Offline, zero deps. |
llmtap (capture) → rewind (replay) → cachelens (optimize)
