# Hermes Agent v0.3.0: token streaming, plugin directory, native Anthropic provider

> Hermes Agent adds token-by-token streaming across the CLI and every gateway platform, a plugin architecture that loads Python files from ~/.hermes/plugins/ without forking, and a direct Anthropic provider that auto-discovers Claude Code credentials instead of routing through OpenRouter. Also lands an ACP server for VS Code/Zed/JetBrains, voice mode, persistent shell state across tool calls, and parallel tool execution — relevant to anyone self-hosting Hermes or scripting against its provider layer.

| | |
|---|---|
| **Tool** | Hermes Agent |
| **Version** | v2026.3.17 |
| **Kind** | release |
| **Published** | 2026-03-17 |
| **Observed** | 2026-08-11 |
| **Significance** | 4/5 |
| **Breaking** | no |
| **Categories** | feature, capability, model-support, security, api-change |

## What changed


- Unified streaming: responses arrive token-by-token in CLI and all gateway platforms instead of as one block (#1538)
- Plugin architecture: drop Python files in `~/.hermes/plugins/` to add tools, commands, hooks (#1544, #1555)
- Native Anthropic provider: direct API calls, Claude Code credential auto-discovery, OAuth PKCE, native prompt caching (#1097)
- Anthropic OAuth: auto-runs `claude setup-token`, reauthentication, PKCE state persistence, identity fingerprinting (#1132, #1360, #1396, #1597)
- Anthropic fixes: adaptive thinking without `budget_tokens` on Claude 4.6 models, cache markers through the adapter, 429/529 retry with surfaced detail, max_tokens/fallback-crash/proxy base_url (#1128, #1216, #1585, #1121)
- Centralized provider router: `call_llm` API, unified `/model` command, provider auto-detect on model switch, direct endpoint overrides for auxiliary/delegation clients (#1003, #1506, #1375)
- Vercel AI Gateway added as a provider (#1628)
- Native Anthropic auxiliary vision instead of OpenAI-compatible routing (#1377)
- Smart approvals that learn which commands are safe, plus `/stop` to kill a running agent (#1543)
- Concurrent tool execution via ThreadPoolExecutor for multi-tool turns (#1152)
- PII redaction before provider calls when `privacy.redact_pii` is set (#1542)
- `/browser connect` attaches browser tools to a live Chrome via CDP (#1549)
- ACP server for VS Code, Zed, JetBrains with slash command support (#1254, #1532)
- Persistent shell mode: cd, env vars, aliases survive across tool calls on local and SSH backends (#1067, #1483)
- Honcho memory: async writes, configurable recall modes, session titles, multi-user isolation in gateway mode (#736)
- Voice mode: CLI push-to-talk, Telegram/Discord voice notes, Discord voice channels, local faster-whisper transcription (#1299, #1185, #1429)
- Agentic On-Policy Distillation training environment added to Atropos (#1149)
- Release notes state 50+ bug fixes across platforms


## Sources

- [github_release](https://github.com/NousResearch/hermes-agent/releases/tag/v2026.3.17) — retrieved 2026-08-11

## Community

_No curated reactions recorded. Facts and community takes are kept in separate layers
and never blended._

---
Canonical: https://changelogs.info/hermes-agent/v2026-3-17
Entity: https://changelogs.info/hermes-agent
Event ID: `evt_2026-03-17_hermes-agent_v2026-3-17`
Licence: event synthesis © changelogs.info, CC BY 4.0. Linked sources belong to their vendors.