Find software your app or agent can use.

One public catalogue record

This is the existing public SpeechBrain example, reviewed 2026-10-16. It is a static sample, supplied without an API request. The full catalogue requires authenticated access under the selected channel’s plan or trial.

{
  "id": "github:speechbrain/speechbrain",
  "name": "SpeechBrain: separate two overlapping voices",
  "purpose": "Separate a two-speaker audio mixture into two replayable WAV tracks.",
  "category": "Audio & recordings",
  "type": "Python speech separation component",
  "platforms": [
    "Python setup component",
    "Local CPU; GPU optional",
    "Separate model download required"
  ],
  "sourceURL": "https://github.com/speechbrain/speechbrain",
  "getSoftwareURL": "https://github.com/speechbrain/speechbrain/blob/c0241278f93ef02cd5191f03caf4e5f8e1292bf6/README.md",
  "licenseDeclaration": "Apache-2.0",
  "licenseURL": "https://github.com/speechbrain/speechbrain/blob/c0241278f93ef02cd5191f03caf4e5f8e1292bf6/LICENSE",
  "requirements": "Setup component, not a finished recorder app. Reviewed SpeechBrain 1.1.2 release/tag c0241278f93ef02cd5191f03caf4e5f8e1292bf6 equals current source (2026-10-14). Python >=3.8.1, matching PyTorch/torchaudio >=2.7.1, soundfile >=0.12.1, HyperPyYAML and other declared dependencies. PyPI publishes the package and matching torch/torchaudio 2.7.1 wheels for Python 3.10-3.13 on macOS arm64. CPU is supported in the actual interface; CUDA is optional. No automatic Apple MPS or speed claim. The ungated Apache-2.0-declared speechbrain/sepformer-wsj02mix checkpoint is a separate first-use download (~113 MB repository storage), not fetched in this review. Narrow route: mono, two-speaker mixture at 8 kHz. It was trained on WSJ0-2Mix; quality on meetings, noisy recordings, accents or more speakers is not established. Output order need not identify the human speakers. Primary model card saves using torchaudio; current SDK also implements soundfile-based audio_io.save. No audio, dependencies or checkpoints installed/executed; source verification only.",
  "evidenceLimits": [
    "Give it a mono recording containing two overlapping speakers and receive two separate 8 kHz WAV files. This supplies separate audio tracks to replay or pass to an existing transcription tool, removing manual attempts to untangle both voices in one mixed track. It separates waveforms; it does not name speakers, transcribe them or establish evidentiary accuracy.",
    "Source review only; software was not installed or run, and individual accessibility fit was not tested."
  ],
  "reviewedAt": "2026-10-16"
}

Direct API

Use your existing direct subscription key. A marketplace key does not work on the direct route. Keep keys in your private client environment.

curl --get "https://www.myworkingmemory.ai/api/software-library/direct/search" --header "Authorization: Bearer $CURATED_FINDS_API_KEY" --data-urlencode "query=SpeechBrain" --data-urlencode "limit=1"

Each admitted search uses one lookup, including a valid empty result. There is no automatic pagination. A cancelled or timed-out request may already have used a lookup; check your allowance before retrying.

query

Optional text, at most 200 characters.

category / platform

Optional existing exact-match filter values, at most 100 characters each.

limit

Integer 1–25. Default 10.

offset

Integer 0–10,000. Default 0. Request each later page explicitly.

The response contains version, total, records, nextOffset and remaining. Compare catalogue versions between pages. Review dates, type labels and missing fields keep their original meanings.

Authentication errors, exhausted allowance and provider failures remain failures. The public client draft sends one request, refuses redirects, bounds responses and makes no startup search. It is not yet published to npm or PyPI.

Direct access currently supports search. It has no direct record-lookup or catalogue-info endpoint. It has no latest, dated archive, date, tag or normalized tool/model filter.

Choose your existing connection

MCPize supplies the existing managed MCP service: search_software, get_software and catalogue_info.

RapidAPI and API.market provide their native REST gateways and document MCP access. Product enablement and a paid MCP response on those gateways still need verification. Apify documents an MCP server for Actors; running this Actor uses its own page billing.

The direct MCP client, Pipedream action, GitHub Action and Python client are prepared distribution drafts. Registry acceptance, client installation, customer credentials and publication remain separate steps. This page does not install or run them.

One product, clear billing units

Request allowances are hard capped. RapidAPI retains its separate 10,240 MB allowance and $0.001/MB bandwidth term. Marketplace trials and entitlements are separate; buying one does not grant access through another.

Direct checkout already uses Stripe. Processing and operating costs apply. Marketplace fees, paid deliveries, refunds, expenses and actual retained profit are measured separately; a listed price is not proof of profit.

Catalogue and product record

2026-10-19: current catalogue has 233 records, retaining each original source-review date.

2026-10-20–21: the current offer is $5.99 per month for 1,000 requests on Direct, MCPize, RapidAPI and API.market. Apify uses completed-page pricing instead of a monthly subscription.

Catalogue version: 6c4567c485c357318c2722eaaaa0d03aae6d6249c663080346bf601d24cae4a0

Product updates

Use this changelog for published product changes. A weekly email edition is being prepared. Email signup and sending will open after the template and collection flow are approved.