whetkit

Measure — then improve — how well LLM agents pick and use the tools your MCP server exposes.

whetkit is a local-first CLI that runs an agent against your MCP server on a set of eval tasks, scores its tool-selection hit-rate, and then curates the tool set — renaming cryptic tools, rewriting vague descriptions, pruning noise, and merging duplicates — via a reversible overlay proxy that never modifies your server. It re-runs the eval through the overlay and hands you a before/after report.

Unlike MCP inspectors and testing frameworks, whetkit closes the optimization loop: it measures agent behavior, proposes a curated tool surface, applies it through a reversible proxy, and re-runs the same evals to quantify the improvement.

whetkit slim      ──►  audit your CLIENT config's union tool bill  (no tasks, no API key)
whetkit doctor    ──►  ten-second lint of one server's surface     (no tasks, no API key)
whetkit inspect   ──►  what does the agent actually see?
whetkit generate  ──►  draft eval tasks from the inventory         (review, then run)
whetkit run       ──►  how often does it pick the right tools?    (hit-rate)
whetkit curate    ──►  fix the tool set, prove it helped           (before → after)

Why tool curation matters

Agents don't read your code — they read your tool names, descriptions, and schemas. A server that grew organically ends up with data_query_1, proc_ord, and do_thing: every one of them costs the model a guess, and every duplicate splits its attention. In practice a large share of agent failures on MCP servers are tool-selection failures — the model calls the wrong tool, loops through near-duplicates, or gives up — and they are fixable without touching a line of server code, because the fix is metadata. whetkit makes that loop measurable: score the failures, patch the metadata through an overlay, and show the hit-rate delta.

Install

uv tool install whetkit   # or: uvx whetkit / pipx install whetkit

(Released to PyPI from tags — see RELEASING.md. PyPI can lag behind main; the quickstart below runs from source.)

Pure Python, no platform-specific code: Linux and macOS are supported and CI-tested on both. Windows is untested — the eval loop spawns stdio subprocesses and --reset-cmd assumes a POSIX shell; reports welcome.

Two-minute start (no API key)

Weigh the tool surface your MCP client actually loads — including servers shipped by Claude Code plugins, which no other audit tool can see:

uv tool install whetkit
whetkit slim --config ~/.claude.json --share

You get tokens + $ per server, cross-server duplicate findings, and a copy-pasteable snippet. Dollar figures are uncached API ceilings — prompt caching and Claude Code's tool search reduce the marginal cost; on subscription plans the real cost is context-window share. Then --recommend-hot --from-traces tells you which servers deserve alwaysLoad: true under tool search, from your real usage.

Quickstart (5 minutes)

Requires uv and an Anthropic API key (or OpenAI — see --model). Python is pinned via .python-version; uv fetches it automatically.

git clone https://github.com/benlamlih/whetkit && cd whetkit
uv sync
export ANTHROPIC_API_KEY=sk-ant-...

1. Inspect the bundled sample server — a deliberately messy e-commerce server (14 tools: cryptic names, vague descriptions, duplicates, noise):

uv run whetkit inspect --server examples/sample-server

2. Baseline eval — run 5 tasks against it and score the hit-rate (deterministic tool-matching + LLM-judge on the final answers):

uv run whetkit run --server examples/sample-server --tasks examples/tasks

3. Curate and prove it — analyze the failures, generate a curation overlay, re-run the eval through it, and get the before/after:

uv run whetkit curate --server examples/sample-server --tasks examples/tasks

This writes:

• .whetkit/curation-plan.yaml — the reviewable, hand-editable overlay plan
• .whetkit/report.html — self-contained before/after report (open it in a browser)
• .whetkit/report.json — the same data, machine-readable
• .whetkit/traces.sqlite3 — full reasoning-path traces of every run

The sample server's failures are tool-selection failures, so weaker models flip several tasks from MISS to HIT through the overlay. Frontier models often ace even the messy baseline on a 14-tool server — there the delta shows up in the other columns instead: tools exposed, tokens per task, and extra calls. On large real-world servers (dozens of tools, near-duplicates) the hit-rate delta comes back.

Stdio servers' own logs are hidden so they can't garble the output; set WHETKIT_SERVER_LOGS=1 to see them when debugging a server that won't start.

The plan is yours to edit: tweak a rename the optimizer got wrong, un-hide a tool, then re-score the curated view directly —

uv run whetkit run --server examples/sample-server --tasks examples/tasks \
  --plan .whetkit/curation-plan.yaml --group curated-v2

4. Use the curated view for real — serve it to any MCP client:

uv run whetkit overlay --server examples/sample-server --plan .whetkit/curation-plan.yaml

Nothing about your origin server changes, ever. Delete the plan and you are back to the original world.

Pointing it at your own server

• --server accepts a URL (streamable HTTP; --http-mode stateless for 2026-07-28-spec servers), a directory containing server.json or server.py, or a .py/.json path directly.
• Write tasks in YAML — format reference in docs/task-format.md — or draft them: whetkit generate --server <your-server> --out tasks/generated.yaml writes candidate tasks from the tool inventory (validated against the live tool list; review before trusting).
• --model / --judge-model / --optimizer-model take provider:model_id, e.g. anthropic:claude-sonnet-5 or openai:gpt-5.2. Keys come from ANTHROPIC_API_KEY / OPENAI_API_KEY.

Commands

Command	What it does
whetkit slim	The scale + the proof for your MCP client config: tokens + $ per message with zero adoption, cross-server dedupe, tool-search-aware --recommend-hot --from-traces (alwaysLoad set from real usage), --share snippet, reversible --apply. Pair with prove-your-gateway.
whetkit doctor	Lint the tool surface: vague descriptions, cryptic names, near-duplicates, context bloat. --json; --fail-on warn for CI.
whetkit inspect	Tool inventory: names, params, description tokens, schema complexity.
whetkit generate	Draft eval tasks from the inventory (server-context aware, read-only unless --allow-writes); review-before-trust YAML.
whetkit run	Agentic eval with real tool execution. --runs N for mean±range and flaky-task detection, --concurrency for independent tasks, --reset-cmd for stateful fixtures, --plan to score a curated view, --summary-json for machine-readable results (with cost estimates).
whetkit curate	Baseline → LLM-proposed overlay plan → curated eval → before/after report. --prune-unused for the cost play.
whetkit fix	Self-correcting curation: propose → eval → feed regressions back → revise, up to --max-iterations; keeps the best plan by measured results.
whetkit plan-init	Scaffold a view plan: --keep a,b, --from-tasks tasks/, --from-traces traces.sqlite3 — hide everything else.
whetkit diff	Compare two --summary-json files: metric deltas + per-task PASS/MISS/FLAKY transitions.
whetkit export	Share a plan: --to markdown (upstream-PR fix table) or --to json (gateway overrides).
whetkit report	Rebuild the HTML/JSON report from stored traces.
whetkit overlay	Serve the curated view as a stdio MCP server.

Authenticated servers: any string in server.json may reference ${ENV_VARS} (e.g. an Authorization header) — credentials never live in the file. Summaries flag probable task-spec gaps, errored runs, and failed tool calls so a bad score always says why.

How scoring works

• Deterministic tool-match: each task lists the expected tool calls (with acceptable alternatives, optionally ordered). Order-tolerant by default, --match-mode exact for strict grading. Reports precision/recall, missing and extra calls.
• LLM-as-judge: grades the agent's final answer against the task's natural-language success_criteria with a strict calibrated prompt; verdicts are cached in SQLite. --judge auto|on|off.
• Hit = right tools and (when judged) task success. The headline metric is the hit-rate across tasks.

More docs: docs/task-format.md · docs/curation.md · docs/prove-your-gateway.md · VERSIONS.md · MIGRATION.md

Development

uv sync
uv run pytest        # full suite, no API key needed (scripted fake provider)
uv run ruff check .

Dependency versions are pinned exactly and documented with sources and check dates in VERSIONS.md. The MCP transport layer supports stdio, legacy stateful streamable-HTTP (2025 spec), and stateless streamable-HTTP (2026-07-28 spec); the SDK-facing code is isolated for the v1→v2 migration (MIGRATION.md).

Scope note: whetkit is deliberately a local-first CLI. Hosting, dashboards, multi-tenancy, and security tooling are out of scope for Stage 1 — the architecture just leaves room for them.

Telemetry

whetkit collects no telemetry by default. If you opt in — with whetkit telemetry on or WHETKIT_TELEMETRY=1 — each command sends one anonymous event containing exactly these fields and nothing else:

• the command name (e.g. run)
• the whetkit version
• the Python version (major.minor)
• the operating system (sys.platform)
• a random anonymous id stored in ~/.whetkit/telemetry.json

Never collected: arguments, file paths, server names, prompts, or results. Sending is fire-and-forget (2 s timeout, failures ignored) and can never block or break a command. Opt out any time with whetkit telemetry off (or WHETKIT_TELEMETRY=0, which overrides the config file); inspect the current state with whetkit telemetry status.

License

Apache-2.0 — see LICENSE.