Hermes Agent v0.7.0 adds pluggable memory, credential pools, and Camofox stealth browser
TL;DR
Hermes Agent v0.7.0 (The Resilience Release) dropped April 3, 2026. It opens memory as an extensible plugin system (Honcho as the reference implementation), adds same-provider credential pools with automatic 401 rotation, integrates the Camofox anti-detection browser, and exposes ACP so VS Code, Zed, and JetBrains can register their MCP servers through the agent.
What changed
What shipped
Hermes Agent v0.7.0 on April 3, 2026. Five engineering-forward changes:
Pluggable memory provider interface
Memory is now an extensible plugin system. Third-party backends (Honcho, vector stores, custom databases) implement a simple abstract base class. Honcho ships as the reference implementation; swap in anything else without forking.
Same-provider credential pools
Configure multiple API keys for the same provider. Hermes uses thread-safe least-used rotation across the pool, and a 401 response triggers automatic rotation to the next key. For production workloads hitting rate limits on a single key, this is a real availability fix.
Camofox stealth browser backend
New anti-detection browser option: stealth browsing with persistent VNC-visible sessions. Useful when you need the agent to interact with sites that fingerprint heavily on headless browser signatures.
Inline diff previews
File write and patch operations now show inline diff previews in the tool activity feed. You can see exactly what the agent is about to change before approving it.
ACP: editor MCP integration
Editor integrations (VS Code, Zed, JetBrains) can now register their own MCP servers, flowing the editor's MCP ecosystem into the agent. If your IDE has MCP servers Hermes did not know about, they show up automatically.
Gateway hardening
The quieter but arguably most important part of the release: race conditions, flood control, stuck sessions, approval routing, compression death spirals all fixed. Secret exfiltration blocking added: the agent will refuse to emit content that looks like an API key or credential in browser URLs or LLM responses.
Availability
Upgrade via standard channel. Honcho remains the default memory backend; no migration needed. Camofox is opt-in via config.
Who this matters for
- Vibe Builder: Inline diff previews show exactly what the agent is about to change before you approve. Credential pools keep the agent running even when one API key hits a rate limit.
- Basic User: Camofox stealth browser and secret exfiltration blocking are under-the-hood wins. Keep using Hermes; it is quietly more reliable.
- Developer: Pluggable memory interface lets you swap in Honcho, vector stores, or custom DBs without forking. Credential pooling with thread-safe rotation is the production availability fix.
What to watch next
v0.7.0's pluggable memory interface is the release detail most people will underestimate. Memory is the single hardest thing about running an agent long-term, and making it swappable (rather than baked in) means Hermes can evolve the memory layer without forcing every user onto the same implementation. If someone builds a better memory backend in 3 months, you can drop it in without forking Hermes.
Credential pools solve the single most common production pain point: rate limits on a single key. Thread-safe rotation plus automatic 401 fallback means you can pool 5 keys for the same provider and get 5x the throughput without writing rotation logic yourself. For anyone running Hermes as a production agent, this is an availability unlock.
Camofox is the feature for a specific but growing use case: agents that need to interact with sites that fingerprint aggressively. Scraping, competitive intel, account management: these all fail on vanilla headless browsers. Persistent VNC-visible sessions also help with debugging when the agent behaves unexpectedly on a specific site.
ACP editor integration is the subtle MCP ecosystem move. Instead of Hermes needing to know about every MCP server, the editor tells it. If your VS Code has MCP servers configured, Hermes sees them. This is how the MCP ecosystem scales without requiring central registration.
The gateway hardening is boring and critical. Any framework at v0.7.0 is going to have race conditions and stuck-session failure modes in production. Cleaning them up before v1.0 is the right order of operations.
by Harsh Desai
About hermes-agent
View the full hermes-agent page →All hermes-agent updatesMore from hermes-agent
- FeatureHermes Agent v0.4.0 adds OpenAI-compatible API, 6 messaging platforms, and @file context
Hermes Agent v0.4.0 (The Platform Expansion Release) shipped March 24, 2026. It exposes Hermes as a drop-in OpenAI-compatible /v1/chat/completions endpoint, adds 6 new messaging adapters (Signal, DingTalk, SMS via Twilio, Mattermost, Matrix, Webhook) bringing the count to 9, introduces @file and @url context injection, adds 4 new inference providers, and enables gateway prompt caching by default.
- FeatureHermes Agent v0.5.0 expands to 400+ models and adds supply-chain hardening
Hermes Agent v0.5.0 (The Hardening Release) shipped March 28, 2026. Nous Portal now routes to 400+ models through one endpoint. Hugging Face becomes a first-class inference provider. Telegram gets private chat topics with per-topic skill binding. Plugin lifecycle hooks activate. Supply-chain hardening removes a compromised dep, pins version ranges, and adds CI scanning.
- FeatureHermes Agent v0.6.0 lands profiles, MCP server mode, and official Docker container
Hermes Agent v0.6.0 (The Multi-Instance Release) shipped March 30, 2026. Profiles let you run multiple isolated Hermes instances from one install. MCP Server Mode exposes the agent's conversations and sessions to Claude Desktop, Cursor, VS Code, and any MCP client. Official Docker container, two new messaging platforms (Feishu/Lark, WeCom), and an ordered fallback provider chain round out the release.