Loading ToolBench
Find the TikTok slideshow formats that work in your niche, with the numbers to back it up. Local MCP server and CLI — your own Chrome, a local SQLite file, no API key.
swipekit has 19 tools with schemas and descriptions present, but quality is inconsistent. Tool names lack clear action verbs (discover, scan_account, top, accounts are vague; only read_slides, save_analysis, write_playbook, track_sound follow verb_noun). Descriptions are brief (10-50 chars) and lack WHEN/WHY context. Parameters have types and descriptions, but many lack constraints (enums, ranges). Output schemas are not documented. Error handling is absent from visible code. The server is STDIO-only, which is a hard cap at 50 for protocol readiness. Overall definition quality is below median due to naming ambiguity and missing output documentation.
the accounts worth copying
search TikTok and collect what comes back (opens Chrome)
everything needed to write a playbook
export the library as a bundle for sharing
which formats recur across UNRELATED accounts
build a production-ready content plan from a playbook
manage product profiles, which tell the agent what you are actually building
read the slides of a post and return them as images
Tool names lack clear action verbs. 'discover', 'top', 'accounts', 'formats', 'sounds', 'stats' do not start with verb_noun pattern (get_, list_, search_, create_). LLMs cannot infer intent from names alone.
Output schemas are not documented. Tools return data but no schema is visible in code for what fields/structure the LLM should expect. This forces LLMs to guess downstream field names and breaks tool chaining.
Descriptions are too brief (10-50 chars) and lack WHEN/WHY context. E.g. 'the posts, ranked by intent rather than views' does not explain when to call top vs search, or what 'intent' means. Descriptions should be 50-200 chars with actionable guidance.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 52 | 2026-07-28+ | v2 |
render an HTML report of the research
start, switch, or list research runs (a run scopes every batch to one question)
save your analysis of a post's format
fetch all posts from a single account (opens Chrome, may take a while)
full-text search posts already collected — instant, no browsing
check whether an account or topic has already been researched, before scraping it again
sounds worth tracking
what's in the library
the posts, ranked by intent rather than views
watch a sound across all accounts using it (opens Chrome)
synthesize a playbook from the evidence pack
No error handling guidance visible. Tools like discover and scan_account open Chrome and may fail (captcha, network, timeout). No recovery hints or error categorization (retryable vs fatal) documented.
Parameters lack constraints. 'sort' enum in top is good, but 'limit', 'target', 'maxFollowers' have no min/max bounds. Unbounded numbers let LLMs pass absurd values (limit=999999, target=1000000).