The 2026-10-10 check used macOS arm64, Python 3.12.13, Ollama 0.40.2, qwen3:4b (359d7dd4bcda), and Laya 0.4.2 from Git commit 68804629e8ccd9d616d48a40e87de9fabbeae069. Laya's English checkpoint revision was 7b928d828b7b0e022f929d9bd2e44165aa270148, running on MPS. The pinned Python environment is requirements-live.txt. These versions were tested on this host; hardware and platform support still depend on the upstream runtimes.
The v0.2.1 release is tagged at eaf8d33fe88374c2a266291def3096737edc18a7. All 14 native/npm/plugin artifacts passed their release SHA-256 checks. The downloaded macOS arm64 executable completed a real MCP stdio handshake with 26 read-only tools, then passed the three-step Qwen/Laya loop with two successful tool calls, dry-run and one fixture update. Existing tags and notes survived; no real Raindrop writes occurred. Homebrew source upgrade, strict formula audit and version/help tests passed locally, and the tap's source build/tests passed in CI. Its source archive SHA-256 is dff61f02b2750565fc7b0092bbb1e6c58e142b51f6793574fbefe3e881d09fee; the installed Homebrew version is 0.2.1.
All seven npm 0.2.1 versions have registry SHA-512 integrity matching the exact GitHub release tarballs. A fresh registry install reports the release commit and 26 read-only tools. The published npx command also completed a real stdio handshake with protocol version 2025-03-26. These checks verify installation, package identity and transport; they do not test every client UI or evaluate model accuracy.
The MCP Registry publication succeeded using short-lived GitHub OIDC. Independent registry API readback confirmed active io.github.KingPsychopath/raindrop-cli version 0.2.1, public title rdrop for Raindrop.io, npm package/version and local stdio transport. Downstream directories control their own indexing; this is a registry listing, not a hosted endpoint or client-directory approval. Listing.
The earlier v0.2.0 release is tagged at c4c997e748655bfa9b615b3d3eb17ea9952cff40. All 14 native/npm/plugin artifacts were downloaded and checked against the release's checksums.txt. All seven published npm versions have registry SHA-512 integrity matching the exact release tarballs. A fresh registry installation reported the release commit and 26 read-only tools. The exact tarballs also installed offline and completed a real MCP stdio handshake. The downloaded native executable also passed the complete three-step Qwen/Laya loop, dry-run and one fixture update described below. This checks package bootstrap and transport independently of model prose. The source archive has its own Homebrew checksum. The tap's source build and formula tests passed in CI and on the tested macOS arm64 host; the installed version at that check was 0.2.0.
To repeat with the released macOS arm64 executable instead of a development build:
gh release download v0.2.1 --repo machineyearningengineer/raindrop-cli \
--dir .artifacts/release-v0.2.1 --clobber
(cd .artifacts/release-v0.2.1 && shasum -a 256 -c checksums.txt)
tar -xzf .artifacts/release-v0.2.1/rdrop_0.2.1_macos_arm64.tar.gz \
-C .artifacts/release-v0.2.1
.artifacts/release-v0.2.1/rdrop_0.2.1_macos_arm64/rdrop version
# Start the model servers below first.
HF_HOME="$PWD/.artifacts/hf" \
LAYA_REVISION=7b928d828b7b0e022f929d9bd2e44165aa270148 \
python3 scripts/verify-models.py \
--binary .artifacts/release-v0.2.1/rdrop_0.2.1_macos_arm64/rdrop \
--output .artifacts/e2e/release-models \
--command-python .venv-models/bin/python
brew test machineyearningengineer/tap/rdropChoose the matching native archive on other platforms. npm uses exact release tarballs and verifies each packaged executable before running it. Check the published version with npx --yes @machineyearn/raindrop-cli@0.2.1 version. Publishing requires npm account 2FA or an eligible granular publishing token. Distribution and publishing commands.
From the checkout:
make build
uv venv --python 3.12 .venv-models
uv pip install --python .venv-models/bin/python -r integrations/laya/requirements-live.txt
# Terminal 1: the first start downloads the pinned checkpoint.
HF_HOME="$PWD/.artifacts/hf" \
LAYA_REVISION=7b928d828b7b0e022f929d9bd2e44165aa270148 \
LAYA_HOST=127.0.0.1 LAYA_PORT=8000 LAYA_MODELS=english \
.venv-models/bin/laya-serveLAYA_DEVICE=cpu selects CPU explicitly. On the tested Apple Silicon host, the runtime selected MPS. Verify the actual device and revision with curl --fail http://127.0.0.1:8000/health. Checkpoints are downloaded model data; no Raindrop credential is required by the local model server.
Install Ollama from its official distribution or Homebrew. Then:
brew install ollama
# Terminal 2:
OLLAMA_HOST=127.0.0.1:11434 OLLAMA_CONTEXT_LENGTH=32768 ollama serve
# Terminal 3:
ollama pull qwen3:4b
ollama listThe pull downloads approximately 2.5 GB. Keep both servers running while verifying; the script does not install or start them. Ollama's Chat Completions compatibility carries the model tool calls; Laya uses the System One decision contract.
HF_HOME="$PWD/.artifacts/hf" \
LAYA_REVISION=7b928d828b7b0e022f929d9bd2e44165aa270148 \
python3 scripts/verify-models.py --command-python .venv-models/bin/pythonThis starts a disposable Raindrop REST fixture on an ephemeral loopback port and supplies a fixture token. It overrides Raindrop configuration for every subprocess. It checks real Laya over HTTP and through the local Python command adapter, then runs the real Qwen model through rdrop agent. The verified run completed in three model steps with two successful tool calls: classification and organization planning. The model received their actual results and returned a response. The captured snapshot plan passes dry-run, then applies only to the fixture. Existing tags and the private note survive; the tag and collection change use one update. The read-only model phase performs zero writes.
Artifacts go to the ignored .artifacts/e2e/models directory: actual tool results, exact plan, dry-run/apply results, fixture requests and summary.json. Repeat with another tool-capable model using --model MODEL --model-url http://127.0.0.1:PORT/v1. The check requires successful calls to raindrop_classify and raindrop_organize_plan; a plausible prose response alone fails it. A different model can fail this task despite implementing the transport.
Real inference exposed interference when two unrelated bookmarks shared one classification state. Requests now isolate each bookmark. After the fix, the API reference received reference (probability 0.9679, confidence 0.8546); the recipe received cooking (probability 0.9319, confidence 0.7306) and stayed marked for review at the default 0.8 confidence threshold. This is a two-example smoke check, not a general accuracy evaluation. Laya warns that some checkpoint temperatures for 11+ choices require clamping; evaluate your own taxonomy, especially larger category sets.
Jev was not contacted because no key was available. Its shared wire contract, malformed responses and uncertainty handling are tested with fixtures. With your key in the private environment, use TYPESAFE_API_KEY and run:
rdrop cleanup report --json
rdrop classify --ids REAL_ID --taxonomy taxonomy.json --jsonJev is a decision provider. A tool-capable model such as Qwen owns the conversational loop and can call raindrop_classify, which delegates to Jev or Laya.
The live official MCP connection discovered all nine documented read tools. Calls to account, semantic bookmark search, extracted content, popular keywords, collections, tags, highlights, misplaced bookmarks and mistagged bookmarks completed without protocol/tool errors. The audit calls selected one tag and a collection containing two bookmarks; empty suggestions are a valid outcome, not an accuracy evaluation. Their private request/result artifacts remain under .artifacts/e2e and are not release assets.
rdrop mcp tools --official --search 'semantic search'
rdrop mcp tools --official --name raindrop_official_find_bookmarks
# arguments.json: {"search":"software development documentation","limit":1}
rdrop mcp call --official --name raindrop_official_find_bookmarks --input arguments.jsonInspect each current upstream schema before preparing other calls. Official coverage records pagination and subscription differences. Fixtures independently prove that upstream mutations and unknown reads are excluded, invalid input stays local, and redirects cannot forward the credential.
The shared plugin skill passed the skill validator. Claude's plugin and marketplace validators passed, and the Codex marketplace resolver recognized raindrop@machineyearn. The pinned MCP launcher still requires the npm package and private CLI configuration; manifest validation alone is not an end-to-end client UI test. Install commands.
The Clef adapter passed fixture tests for the Workers AI request/envelope, isolated bookmark state, review thresholds and invalid account/model paths. No Workers AI credentials were configured for live Clef inference. Jev/Clef live quality and local Clef serving remain separate verification work.
rdrop cleanup report --json > cleanup-report.json
RDROP_DECISION_BASE_URL=http://127.0.0.1:8000/v1 \
RDROP_DECISION_MODEL=english \
rdrop agent --base-url http://127.0.0.1:11434/v1 --api chat --model qwen3:4b --json \
'/no_think Review selected bookmarks, classify them using existing categories, and propose organization. Do not apply changes.' > agent-result.json
jq '.plans[0] // error("No accepted plan")' agent-result.json > reviewed-plan.json
# Inspect the actual tool results, uncertainty and exact plan before these steps.
rdrop cleanup apply reviewed-plan.json --dry-run
rdrop cleanup apply reviewed-plan.json --yesThe last command intentionally changes the selected bookmarks. Full contexts and plans can include private metadata; keep these files private. Omit the apply commands when only reviewing.
The tested local host is executor@2.0.0-beta.14, reporting apps@0.0.1-beta.69. The published npm beta was checked directly; the website described beta.15 before that package was available. Use the host's actual framework version rather than assuming a website version has been released.
See the integration guide for installation, source preparation, creation and deployment. The local wrapper uses ~/.local/share/rdrop-executor/keys.json; the host keeps that file private and stores connected account credentials encrypted. OS-backed key storage is also supported by Executor; the wrapper is specifically for this file-backed local setup. Never commit that directory or copy keys into MCP configuration committed to Git.
After deployment, create a profile and issue an account connection through Executor's management tools, using the exact paths returned by tools.search. The secure form selects the account for that profile. Run the provider health check and read cleanup.report, tags and collections through the deployed profile before using writes. The adapter's two mutation tools still require exact-call host approval; deployment and account connection do not grant blanket mutation approval.
The real local deployment passed the provider account health check and read cleanup counts, tags and collections through its MCP profile. No live bookmark mutation was performed.
For reproducible CLI calls, python3 scripts/executor-mcp.py lists MCP tools. Save an execute request as a private JSON file:
{"code":"return await tools.search({namespace:'executor',query:'appProfiles create accountConnect issue accounts check',limit:5});"}python3 scripts/executor-mcp.py --tool execute --input PRIVATE_REQUEST.jsonInspect the returned signatures, then use the same command with the actual management calls. A call can require approval or input: show the question to the user and resume that request; never rerun a paused program or approve on their behalf. This helper uses the local host at http://127.0.0.1:4312/mcp; hosted setup has a separate authentication and organization boundary.