Hermes Agent 0.18 and OpenClaw v2026.7 updates bring coding projects and mobile agents

What shipped

From 13 to 20 July both Hermes Agent and OpenClaw shipped repeated infrastructure and feature updates focused on agent reliability and coding workflows. Hermes Agent stabilized its gateway and added verification contracts while OpenClaw expanded to mobile and introduced guided setups. The releases center on practical agent tooling rather than new base models.

Hermes Agent

Hermes Agent delivered its v0.18 series with a focus on coding projects inside the desktop app and self-verification loops. The updates also improved gateway scaling, Google Vertex AI integration, and security hardening across multiple patch releases. Builders now get concrete tools for managing skills and verifying task completion against real evidence.

• •Gateway scale-to-zero Hermes-agent now supports idle dormancy and external drain coordination so hosted setups can scale to zero without losing active conversations.
• •Google Vertex AI support The framework added first-class Gemini model access with automatic OAuth2 token minting and refresh from GCP service accounts.
• •Desktop coding cockpit A dedicated Projects view appeared in the desktop app complete with sidebar, coding rail, review pane, git worktrees, and multi-terminal panels.
• •Skill commands in CLI The new /learn command distills reusable skills from folders or URLs while /journey lets users view and edit skill timelines.
• •Completion contracts /goal loops now use contracts and a verification evidence ledger so the agent checks work against actual project outputs instead of model claims.
• •Mixture-of-Agents picker MoA ensembles became selectable in the model picker with live streaming of individual model reasoning and final aggregated answers.
• •Windows installer fixes Version 0.18.1 rolled up 660 pull requests including self-healing installers, dashboard fixes, and WhatsApp pairing stability.
• •WhatsApp dependency fix The Baileys library was unpinned to the published 7.0.0-rc13 release to restore reliable Docker builds.
• •Security hardening MCP config persistence was locked down, cron-based credential leaks were blocked, and Slack tokens are now redacted.
• •Vertex AI token refresh Automatic short-lived OAuth2 token handling prevents mid-session expiry when using Google Vertex AI models.
• •Gateway drain coordination The gateway now scales to zero on idle and coordinates external drains to avoid dropping conversations during updates.
• •Dedicated coding projects Projects received a sidebar, coding rail, review pane, and git worktree management inside the desktop app.
• •Background subagent delegation delegate_task can now fan out multiple subagents while the user continues chatting and receives a consolidated summary later.
• •Skill distillation commands /learn extracts reusable skills from directories or URLs and /journey displays an editable timeline of memories and skills.
• •Self-verification for goals Completion contracts and a coding evidence ledger let the agent prove work against real checks rather than stopping early.
• •MoA model selection Users can pick Mixture-of-Agents presets and watch individual reasoning blocks stream before the final answer appears.
• •v0.18.1 stability rollup Over 660 pull requests were merged with Windows installer self-healing and gateway fixes included.
• •Docker build reliability Switching to the published Baileys npm package resolved tagged-release Docker image failures.
• •MCP and credential security Locked MCP config files and blocked exfiltration paths via cron overrides were added.
• •Native Vertex AI provider Google Vertex AI joined as a first-class provider with automatic token refresh from service account JSONs.
• •Idle gateway scaling The gateway now scales to zero when idle and coordinates drains before restarts.
• •First-class coding projects A Projects sidebar with coding rail and git worktree support was added to the desktop app.
• •Skill timeline commands /learn distills skills from URLs or folders while /journey lets users manage the accumulated timeline.
• •Goal verification ledger Completion contracts now require evidence from actual project checks before marking tasks done.
• •Ensemble model routing Mixture-of-Agents presets were integrated into the picker with live reasoning block streaming.
• •Infrastructure stability patch 660 pull requests delivered Windows installer fixes and MCP provider improvements.
• •Baileys dependency update The library was switched to the published npm release to fix Docker build issues.
• •v0.18.0 major release Mixture-of-Agents, self-verification, background subagents, and desktop coding Projects all landed together.
• •v0.18.1 patch release Windows installer self-healing and gateway fixes were consolidated from 660 pull requests.
• •v0.18.2 dependency patch The WhatsApp Baileys library was updated to the published npm version to restore Docker builds.

OpenClaw

OpenClaw focused on mobile parity and easier onboarding with new Linux bundles, offline caching, and guided Control UI setup. The week also brought expanded model support including GPT-5.6 defaults and remote coding session handling. Builders gained concrete improvements in channel safety and automation reach across platforms.

• •Linux install bundles New deb and AppImage packages include Gateway guidance and Windows winget installs now continue after Node.js setup.
• •iOS offline chat Recent sessions and transcripts are pre-painted from a bounded gateway cache while sending stays disabled offline.
• •GPT-5.6 default model Fresh API and Codex setups now default to openai/gpt-5.6 with medium reasoning preset.
• •Guided provider setup Settings now let users configure model providers and onboard channels through step-by-step pages.
• •Channel safety fixes Telegram ingress loss after restarts was prevented and Signal controls stay responsive during turns.
• •Mobile automation parity Automations, Voice Wake on Android, and camera or location access from headless Linux nodes were added.
• •Remote coding sessions Control UI sessions can run on cloud workers while Codex and Claude sessions open in owning-host terminals.
• •ClickClack CLI setup Interactive prompts configure ClickClack from the CLI and bot commands appear in composer autocomplete.
• •External gateway mode OPENCLAW_SUPERVISOR_MODE=external preserves restart behavior for external lifecycle owners.
• •Linux packaging updates Stable main-based deb and AppImage bundles were published with improved Windows winget handling.
• •Offline session caching iOS apps now pre-paint sessions from a protected per-gateway cache when offline.
• •Default model change openai/gpt-5.6 becomes the default for new API-key and Codex setups with medium reasoning.
• •Provider onboarding flow Settings and a guided page handle model provider configuration and channel selection during session creation.
• •Native mobile automations Automations parity reached mobile with foreground Voice Wake on Android and Linux node capabilities exposed.
• •Terminal session resume Codex and Claude catalog sessions can open directly in terminals on their owning hosts.
• •Worktree cleanup limits Settings now let users set count and size limits for managed worktree cleanup.
• •MCP session scoping Model Context Protocol server connections are now limited to the requesting session only.
• •Guided channel setup A step-by-step page helps onboard channels and pick images or models at session start.
• •Headless Linux features Camera, location, and notification access were exposed from headless Linux nodes.
• •Cloud worker sessions Control UI sessions can run on cloud workers with terminal resume for OpenCode and Pi sessions.
• •Gateway crash loop fix Repeatedly failing gateways now leave a stable repair path instead of indefinite restarts.
• •Messaging channel upgrades Telegram, Slack, Discord, and Apple Messages gained live progress, media handling, and voice sessions.
• •Codex workflow tools The openclaw attach command grants temporary access and delegation to connected coding agents was improved.
• •New model additions GPT-5.6, Tencent Hy3, and Meta Muse Spark 1.1 joined the supported provider list.
• •Mobile app updates Official iOS, Android, and macOS apps received improvements to navigation, voice, offline reading, and session controls.
• •Control UI redesign Side-by-side conversations are easier with live Tasks, clearer controls, and usage or cost views kept nearby.
• •v2026.7.1 release New providers including Featherless, Claude Sonnet 5, and Mythos 5 were added.
• •Standalone Meta provider A new package bundles Muse Spark 1.1 Responses API support with streaming and tool calls.
• •Container migration safety Migrations now run before Gateway readiness and a safe mode prevents restart flapping after unclean boots.
• •Telegram steering commands Private login pairing and new steering commands were added for active Codex runs.
• •Automatic session titles Utility-model routing now generates concise titles and Gateway-backed groups help organize conversations.
• •Mobile offline caching Bounded sessions and transcripts are cached on iOS and Android with Apple Watch voice turns supported.
• •Crestodian onboarding An agent-loop setup runs across CLI, web installer, and macOS app with masked credential prompts.
• •Control UI overhaul A minimal searchable sidebar and reasoning-effort slider were added alongside a native macOS session browser.
• •Expanded model list Featherless, Claude Sonnet 5, Mythos 5, and GPT-5.6 are now available with optimized reasoning presets.
• •Beta release v2026.7.2-beta.3 Infrastructure updates continue in the latest beta.
• •Beta release v2026.7.2-beta.2 Further stability work landed in this beta.
• •Beta release v2026.7.2-beta.1 Initial beta for the 7.2 series was published.

What this means for you

For Vibe Builders: You can now run coding agents with desktop projects and self-verification in Hermes Agent while OpenClaw gives mobile automations and offline chat without writing code. Start with the guided Control UI setup or the new /learn command to build reusable skills quickly. Both tools let you test agent workflows on your own machine this week.

For Non-techies: Hermes Agent and OpenClaw now handle more daily tasks like offline messaging and automatic model selection so you spend less time on setup. Install the Linux bundles or iOS updates to keep chats running when connectivity drops. The guided provider screens make adding new models straightforward for small business use.

For Developers: Hermes Agent v0.18 added completion contracts and Mixture-of-Agents routing while OpenClaw improved MCP scoping and gateway supervisor mode. Evaluate the new gateway drain coordination and worktree cleanup limits against your existing agent runtimes. Watch the next beta for further Codex delegation changes before integrating into production pipelines.

What to watch next

Track Hermes Agent v0.19 for further verification ledger expansions and OpenClaw v2026.7.3 for additional remote terminal controls. Monitor MCP provider stability metrics and iOS offline cache limits in the coming releases.

Sources

• •[v2026.7.2-beta.3](https://github.com/openclaw/openclaw/releases/tag/v2026.7.2-beta.3)
• •[v2026.7.2-beta.2](https://github.com/openclaw/openclaw/releases/tag/v2026.7.2-beta.2)
• •[v2026.7.2-beta.1](https://github.com/openclaw/openclaw/releases/tag/v2026.7.2-beta.1)