The tool name shortening logic was correctly finding the tool in the
database but then passing the reconstructed parsed.toolName (which may
have an incorrect 'Tool' suffix) to the executor instead of using the
actual tool name from the database record.
This caused 'tool not found' errors when executing tools via MCP SSE
even though the tools were listed correctly.
- Updated deleteApiKey description from 'Delete an API key.' (18 chars)
to 'Delete an existing API key from the account.' (44 chars)
- Bumped version to 0.1.1
- Add parameters to tool API responses (agents + collections)
- Update ToolInfo interface to include parameters
- Render parameters grouped by required/optional
- Show parameter name, type, description, and default value
- Style with design system patterns (fieldsets, badges, etc.)
- Add GPT-4.1 and GPT-4.1-mini to OpenAI provider models
- Set gpt-4.1-mini as default model for new agents
- Make collections in chat tools panel expandable/clickable
- Fix type error in ChatToolsPanel
AI SDK requires every tool call to have a matching tool result.
When a tool call's result is missing (e.g., due to network error or
interrupted execution), the conversation would break and users couldn't
send new messages.
This fix:
- Builds a set of tool call IDs that have corresponding tool results
- Only includes tool calls in assistant messages that have matching results
- Falls back to text-only content if all tool calls are missing results
- Pass excludeIds param to search API to filter out already-added tools
- Increase search limit to 50 (was 10) for large packages
- Increase max API limit to 100 (was 50)
- Add notIn filter to database query for efficient exclusion
- Add package name match boost to search ranking (50 points for package name matches)
- Keep dropdown open when adding tools for faster bulk additions
- Close dropdown only when all results have been added
- Replace minimal tabs with comprehensive API documentation
- Add fieldset-style containers with dashed borders per design system
- Document all 4 endpoints: send message, list, get, delete conversations
- Add authentication section with clear instructions
- Add parameter tables with types, defaults, and descriptions
- Document SSE event types for streaming responses
- Add conversation ID explanation section
- Use lowercase headings and monospace fonts per design guide
- Code examples in cURL, TypeScript, and Python
- Always show McpUrlSection on public collection pages (not just owners)
- Add note for non-owners about providing their own credentials
- Add API usage example for non-owners on collections
- Add AgentApiSection component for public agent pages
- Show conversation API endpoint and usage example
- Make Chat button available to everyone on public agent pages
Users can now access other users' PUBLIC agents and collections by providing
their own credentials in the request:
For agents:
- Provide `providerApiKey` for LLM access
- Provide `env` object with tool environment variables
- Owner's stored credentials are never shared
For collections (MCP):
- Provide `env` in params for tool environment variables
- Owner's stored credentials are never shared
Returns clear errors listing missing required env vars if not provided.
Files changed:
- packages/types/src/agent.ts: Add providerApiKey to SendMessageSchema
- apps/web/src/lib/agents/env-helpers.ts: New helper functions for env vars
- apps/web/src/lib/agents/build-tools.ts: Accept callerEnvVars parameter
- apps/web/src/app/api/agents/[id]/conversation/[conversationId]/route.ts:
Allow public agent access with caller credentials
- apps/web/src/lib/mcp/handlers.ts: Accept callerEnvVars, validate requirements
- apps/web/src/app/api/mcp/[username]/[slug]/[transport]/route.ts:
Allow public collection access, pass isOwner flag
- apps/web/src/app/docs/platform-guide/page.tsx: Update access model docs
Document how users can view the parent agent/collection to see:
- Current tools and collections the parent has
- How to manually update fork with parent changes
- API endpoint for fetching parent data programmatically
Add comprehensive documentation explaining:
- Fork-to-use access model for agents and collections
- API access behavior for agents (403 errors, owner pays for LLM)
- MCP access behavior for collections (403 errors, caller pays)
- Common access error codes (401, 403, 400, 429)
- Environment variable and API key requirements
- Fix Drawer, Modal, Popover, Tooltip, DropdownMenu imports and usage
- Use controlled component pattern with open/onClose props
- Fix Breadcrumbs to use BreadcrumbItem children instead of items prop
- Fix Slider onChange to handle ChangeEvent properly
- Fix Pagination prop name from currentPage to page
- Fix EmptyState icon from invalid "inbox" to valid "box"
- Fix StatCard props to use value (number), label, and subtext
- Add SWR package and global SWRProvider
- Create reusable hooks: useTools, useAgents, useCollections,
useStats, useLikeStatus, useBundleSize, useActivity
- Update tool-search page to use useTools hook
- Update agents page to use useAgents hook
- Update BundleSize component to use useBundleSize hook
- Update LikeButton with optimistic updates via useLikeStatus
- Add simple about page with creator info
- Add sticky left-hand navigation with scroll-spy for active section
- Add 20+ missing UI components to style guide showcase
- Include Accordion, Breadcrumbs, Drawer, Modal, Pagination, etc.
- Add state components: EmptyState, ErrorState, LoadingState
- Remove Playground link from header and mobile menu
- Comment out Architecture Diagram on homepage temporarily
- changelog: Replace puzzle and github SVGs with Icon, use Button for CTA
- sdk: Replace 4 github SVGs with Icon component
- Use design system tokens for version badge colors
- Replace raw <label> with Label component in auth pages, profile settings, tool-ideas
- Replace raw <button> with Button in CopyButton, CopyDropdown, ExecutorConfigPanel, PackageManagerSelector, ToolPlayground
- Replace raw <textarea> with Textarea in ToolPlayground
- Replace inline SVG with Icon component in error.tsx
The rollup-plugin-dts used by tsup runs out of memory with 54 entry points.
Split build into two steps:
1. tsup for ESM bundles
2. tsc --emitDeclarationOnly for .d.ts files
Add new reusable components:
- EmptyState: for empty list/search states with icon, title, description
- ErrorState: for error displays with optional retry button
- LoadingState: for loading states with Spinner and message
- PageHeader: for consistent page headers with title, description, actions
Apply components across pages:
- agents, collections, tool-search pages use new state components
- stats page uses LoadingState and ErrorState
- faq page uses Icon for chevrons
- health page uses Table components
- dashboard agent forms use Input, Select, Textarea, Label
- Replace custom div spinners with Spinner component
- Replace hardcoded color classes with design system tokens (error, success, warning, info)
- Replace raw HTML inputs with Input component from @tpmjs/ui
- Replace raw HTML buttons with Button component
- Add eye/eyeOff icons for password visibility toggle
- Add mail icon for verify-email page
- Fix text colors to use semantic tokens (text-error, text-success, text-warning)
- Fix background colors to use opacity tokens (bg-error/10, bg-success/10)
- All pages now properly support dark mode through CSS custom properties
New UI Components:
- Modal/Dialog with sizes, focus trap, backdrop
- Toast/Notification with variants, stacking, actions
- Drawer/Sheet with directions and widths
- Popover with triggers and positioning
- Tooltip with delays and placements
- DropdownMenu with items, dividers, keyboard nav
- Breadcrumbs with separators and collapsing
- Pagination with full/simple/minimal variants
- Accordion with single/multi expand modes
- Skeleton with text/avatar/card/table variants
- InstallSnippet with package manager toggle
- QualityScore with tier badges
- ToolCard for registry display
Design System Enhancements:
- Complete token specification (shadows, radius, z-index, opacity)
- Full dark mode palette in globals.css
- Updated component variants for dark mode support
Style Guide Expansion:
- Split into 21 modular section files
- Pattern library: navigation, forms, feedback, tables, search
- Governance: a11y checklists, content guidelines, icon system
- Interactive examples throughout
Claude Desktop enforces a 64 character limit on MCP tool names.
- Remove 'tpmjs-tools-' and 'tpmjs-' prefixes from package names
- Remove 'Tool' suffix from tool names
- Add fallback truncation if still too long
- Update parser to try all possible package/tool name combinations
The test was using ctx.auth.username from env vars, which may not match
the actual username of the API key owner. Now fetches the real username
via /api/user/profile endpoint before running MCP tests.
Add "Example Use Cases" section to collection pages that generates
practical workflow examples showing how tools can work together.
- Add useCases and useCasesGeneratedAt fields to Collection model
- Create use-cases-generator.ts using Vercel AI SDK with gpt-4.1-mini
- Add POST /api/collections/[id]/use-cases/generate endpoint
- Add AI_GENERATION_RATE_LIMIT (5 req/hour per IP)
- Create UseCasesSection component with generate/regenerate UI
- Generate 6 use cases: 3 simple (1-2 tools) + 3 complex (3-5 tools)
- Include useCases in public collection API response
Complete unsandbox API coverage with 7 total tools:
- executeCodeAsync: async code execution
- execute: sync code execution
- run: sync with shebang auto-detect
- runAsync: async with shebang auto-detect
- getJob: get job status/results
- listJobs: list all active jobs
- deleteJob: cancel a job
