release aidev BREAKING SIG 4/5

Hermes Agent v0.4.0: OpenAI-compatible API server, 6 messaging adapters, prompt caching

Hermes gains an OpenAI-compatible /v1/chat/completions endpoint plus a jobs REST API, six new messaging platform adapters, and four new inference providers. Gateway prompt caching preserves the Anthropic cache across turns, and gateway approval moves from bare text to explicit /approve and /deny commands.

PUBLISHED2026-03-24
OBSERVED2026-08-11
AGE4mo
SOURCES1
  • OpenAI-compatible API server exposing /v1/chat/completions, plus /api/jobs REST API for cron management; hardened with input limits, field whitelists, SQLite response persistence, and CORS origin protection
  • New adapters: Signal, DingTalk, SMS (Twilio), Mattermost, Matrix, Webhook — joining Telegram, Discord, WhatsApp; gateway auto-reconnects failed platforms with exponential backoff
  • New inference providers: GitHub Copilot (OAuth + token validation), Alibaba Cloud/DashScope, Kilo Code, OpenCode Zen/Go
  • @file and @url context references with CLI tab completion
  • hermes mcp CLI for installing, configuring, and authenticating MCP servers over OAuth 2.1 PKCE
  • Gateway prompt caching: AIAgent instances cached per session so the Anthropic prompt cache survives across turns, cutting cost on long conversations
  • Context compression overhaul — structured summaries with iterative updates, token-budget tail protection, configurable summary endpoint, fallback model
  • Behaviour change: CLI streaming is now on by default, with spinner and tool-progress display during streaming
  • Behaviour change: gateway approval requires explicit /approve and /deny commands, replacing bare text approval
  • New CLI commands: /statusbar, /queue, /permission, /browser, /cost
  • 200+ bug fixes, including streaming chunk concatenation and linebreak handling, and spinner suppression in non-TTY environments
featureapi-changebreakingcapabilitymodel-supportpricing

COMMUNITY

No curated reactions recorded for this event. Facts and takes are kept in separate layers — community context is added by hand, never blended into the record above.