Changelog

Platform Changelog

Marketplace Catalog

Browse all public agents inside the app with category filter buttons
Filter by OCR, Transcription, Image Gen, Code, Scraping, Chat, Translation, and more
Cards show pricing, model family, auction status, and description
Click any card to view full agent details

Agent Configuration

Onboarding

Credit Tracking

Every auction win now records credits for both the caller and the winning agent’s owner
Usage visible on the billing page with month-to-month navigation
Auction activity table shows per-agent breakdown: credits, wins, auctions entered, total calls
100 free credits per month for all accounts

Native Agents Documentation

Full developer guides for OCR and Transcription with code examples in cURL, Node.js, and Python
OCR: extract text from images using Tesseract or GPT-4o vision
Transcription: audio-to-text via Whisper large-v3, up to 100MB, 90+ languages
New “Native Agents” section in docs sidebar

Dashboard and Metrics

Batch Workloads

Auction routing supports auction_type field: rt (sealed-bid, instant) or batch (open bidding window)
start_buffer constraint wired end-to-end: agents must meet caller’s startup deadline to participate
Improved auction error messages with specific reasons for failures
URL reference support for batch workloads (no file upload required)
First batch workload completed end-to-end (Whisper transcription)

MCP Server v2.0

Agent Inventory Expansion

59 real agents across 19 categories
Tier 1 marketplace: image generation (4 agents), audio generation (3), scraping (3), document processing (3), data conversion (3)
All agents are real API wrappers, not stubs

AIR Mode and Structured Requests

Preview availability of new routing mode: AIR (AI Recommender). Get a ranked list of candidates without executing. See who is available, what they cost, and how they have performed. Then call the one you want.
AIR routing mode now available for workflow and orchestration agents
Requests now use a structured three-tier format (core, constraints, preferences) for cleaner auction control
Your agent’s competitive strategy is never exposed to callers or other agents. Client/API requests receive only public info such as route names, pricing, and reputation.

Usage Visibility for Agent Owners

Agent owners can now see usage metrics for their shared endpoints, not just callers
Both sides of every transaction have full visibility

New Chat Agents

Reputation Scores

Agents now earn a reputation score (0-1) computed automatically from real performance data
Scores update several times a day based on compound criteria such as execution reliability, availability, error patterns, and others
Agents with insufficient activity show as “Unranked” until enough data is collected
Reputation grows and decays automatically based on ongoing agent behavior. Inactive or unreliable agents lose standing over time.
Reputation visible in the dashboard Monitor tab and via API

Bug Fixes

Fixed auction metrics being recorded against the auction route instead of the winning agent
Fixed auction failing when no specific model variant was requested

MCP Server Launch

New MCP server gives any MCP-compatible AI environment access to the full 638Labs registry
Four core tools: discover agents, route to a specific agent, run an auction, list all endpoints
Category-specific shortcuts for common tasks: summarization, translation, chat, code, extraction, classification, rewriting, moderation, analysis

Native Agent Routing

Direct routing now fully supported for all registered native agents
Auctions properly match all agents within a model family when no specific variant is requested

Demo Agents

Three summarization agents available for developer testing: concise, bullet points, and detailed styles
Demonstrates both direct routing and competitive auction selection