Verdict: Hermes Agent's two biggest mid-2026 upgrades — a session-database optimization that shrinks on-disk storage by up to 78% and an optional tldraw offline skill that lets your agent script live interactive whiteboards — are free, shipped in recent releases, and genuinely change what your agent can do. The database fix just runs itself; the tldraw skill turns your agent from a text-only tool into one that can think visually and build small interactive apps on a canvas. Combined with the broader Hermes ecosystem (Kanban multi-agent swarms, HTML-to-video Hyperframes, and subagent delegation), a single free open-source agent now handles work that used to require a stack of paid tools.
Last verified: 2026-07-30
- Session DB optimization: up to 78% disk reduction, ~60% average — vendor-reported (source)
- tldraw offline skill: optional install via
hermes skills install official/creative/tldraw-offline(docs)- Kanban multi-agent board: a first-party Hermes feature since v0.15 (docs)
- Hyperframes: optional HTML-to-video skill (docs)
- All free and open-source (MIT license)
- Pricing/availability note: features ship in recent Hermes releases (v0.18+ as of July 2026). Run
hermes updateto get them.
What Is the Hermes Agent Session Database Optimization?
Hermes Agent saves every conversation — from the CLI, Telegram, Discord, Slack, WhatsApp, Signal, Matrix, Teams, and other platforms — as a session in a local SQLite database at ~/.hermes/state.db (Hermes docs — Sessions). This database uses WAL mode for concurrent reads and includes FTS5 full-text search so you can search across past conversations. As you use Hermes throughout the day, this file grows continuously.
On July 22, 2026, Nous Research announced a new optimization that reduces the on-disk size of this database by up to 78%, with an average reduction of approximately 60% (official X announcement). These figures are vendor-reported — no independent benchmark has been published as of this writing (The Agent Times, July 22, 2026).
How it works
The optimization has two rollout paths:
| Database size | What happens | User action required |
|---|---|---|
| Under 1 GB | Transition triggers automatically on next update | None — it runs silently in the background |
| Over 1 GB | Manual command needed | Run hermes sessions optimize-storage once after updating |
Key design property: your search stays live the entire time the optimization runs. The process works in the background without taking any functionality offline, so you can keep searching past sessions while the database shrinks.
A separate GitHub issue (#43690) documented the root cause behind some database bloat: the FTS5 trigram index was expanding JSON tool-call data by up to 18.3×, inflating a 27 MB payload to 502 MB of trigram entries. The fix involves excluding structured JSON from trigram indexing and using detail='none' to reduce FTS5 overhead by ~37.5% (GitHub Issue #43690). This is the engineering work behind the headline 78% number.
What this means for you
If you run Hermes daily — especially across multiple platforms or with multiple agent profiles — your state.db file was getting measurably heavier over time. After this optimization, all that accumulated session history takes a fraction of the disk space it used to while remaining fully searchable. For cloud deployments where disk space costs money, this is a direct cost reduction. For local installations on laptops or small VPS instances, it means Hermes stays lightweight no matter how long you use it.
How Does the Tldraw Offline Whiteboard Skill Work?
The tldraw offline skill is an optional Hermes skill that lets your agent drive the tldraw offline desktop app — building diagrams, wireframes, layouts, and flowcharts on a real canvas by writing code, not by clicking around a GUI (Hermes docs — Tldraw Offline).
The core insight: scripting beats drawing
Here is what makes this skill different from any other "AI draws on a whiteboard" demo: Hermes does not drag boxes around the canvas like a person would. Instead, it writes JavaScript document scripts — code that generates the shapes programmatically. An agent is far better at scripting a canvas than at clicking around on one.
The tldraw offline app runs a local HTTP API (default localhost:7236) that Hermes drives with plain curl from the terminal. Every request requires a bearer token from server.json in the app's config directory. This is the same mechanism the app's own homepage demo uses.
What you can build
| Capability | Example |
|---|---|
| Static diagrams | Architecture diagrams, flowcharts, wireframes, layouts |
| Reactive shapes | Shapes that respond to pointer events and interactions |
| Interactive buttons | Clickable buttons driving state changes on the canvas |
| Counters and state | A number display with +/- buttons that actually work |
| Connection logic | Shapes that react when connected or touched |
The board is not just a picture — it can behave like a small interactive app. A regular whiteboard cannot do that. The tldraw skill ships with a counter example (scripts/counter.js) that demonstrates a number display with MINUS, RESET, and PLUS buttons, all driven by pointer events on drawn shapes.
Step-by-step: set up the tldraw offline skill
Install the tldraw offline desktop app from the GitHub releases page (available for macOS, Windows x64/Arm64, Linux x86_64/arm64 as AppImage or .deb).
Install agent skills in the app: In tldraw offline, go to
Develop → Install Agent Skills. The app writes its skill files into the Hermes skills directory.Install the Hermes skill: Run
hermes skills install official/creative/tldraw-offline.Open a document in tldraw offline and keep the app visible on your screen. Hermes needs the app running to reach the local API.
Ask Hermes to build something: Describe the diagram, layout, or interactive board you want. Hermes reads the current token/port from
server.json, then uses the local HTTP API to script the canvas.
Critical pitfalls to avoid
- Keep tldraw open: If the app is not running with a document visible, Hermes cannot reach the board and every API call fails.
- Re-read port + token on every shell call: The port can be non-default (
server.listen(0)picks one dynamically). Always readserver.json; never hardcode7236. - Scripts are idempotent: Use
createShapeIfMissingwith stable IDs, or you will duplicate content and clobber user edits on every script rerun. - Clean up on signal: Every
store.listen/editor.on/setIntervalmust attach cleanup to theAbortSignal, or listeners pile up across reruns.
What Is the Hermes Kanban Multi-Agent Swarm?
Beyond the two headline upgrades, Hermes ships a Kanban multi-agent board — a durable task board backed by a SQLite database at ~/.hermes/kanban.db that lets multiple named agent profiles collaborate on work (Hermes docs — Kanban).
This is not the same as delegate_task (which is a function call inside a running agent's session). Kanban is durable: every task is a row in the database, every handoff is persisted, and workers run as full OS processes with their own identity and workspace. The board survives restarts, crashes, and host reboots.
How a Kanban swarm works
The Kanban swarm topology helper creates a graph in one shot:
hermes kanban swarm "Design a multi-region failover plan" \
--workers researcher,architect,sre \
--verifier reviewer \
--synthesizer writer
This creates:
- N parallel worker cards — each assigned to a specialist agent profile
- A verifier card — gated on all workers completing (checks their output before the synthesizer sees it)
- A synthesizer card — gated on the verifier marking the work clean (combines verified output into a final deliverable)
Workers run in parallel; the board tracks state so a crashed or restarted worker simply picks up its last in-progress card. You can watch the whole thing unfold on the Hermes dashboard and jump in with comments at any time. The /kanban slash command works from any messaging gateway — Telegram, Discord, Slack, WhatsApp, Signal, Matrix, and others — so a chat message can spin up a full swarm without touching the CLI.
The verifier pattern is the moat: a swarm without a verifier is just delegate_task with extra steps. The verifier checks specialist output before the synthesizer sees it — that is what makes the result trustworthy (awesome-hermes-usecases).
Kanban vs. delegate_task: when to use which
| Factor | Kanban | delegate_task |
|---|---|---|
| Durability | Survives restarts and crashes | Only lives during the parent's session |
| Cross-agent handoffs | Yes — any profile can pick up a card | No — parent-child only |
| Human input | Yes — cards can be blocked for human review | No — subagents cannot clarify |
| Discoverability | Board is queryable after the fact | Results enter parent context, then lost |
| Best for | Long-running multi-agent work | Quick parallel reasoning inside one conversation |
For more on building agent fleets on Hermes, see our walkthrough: How to Build an AI Agent Fleet on Hermes Agent 0.19: The Small-Business Walkthrough.
How Does Hyperframes Turn HTML Into Video?
Hyperframes is another optional Hermes skill that renders MP4/WebM video files from HTML compositions (Hermes docs — Hyperframes). The principle: HTML is the source of truth for video. A composition is an HTML file with data-* attributes for timing, a GSAP timeline for animation, and CSS for appearance.
npx hyperframes init my-video # scaffold a project
npx hyperframes preview # live-reload browser preview
npx hyperframes render --output final.mp4 # render to MP4
This lets your agent generate explainer videos, social overlays, captioned narration, audio-reactive visuals, and scene transitions — all locally on your machine, without a paid video tool. You can even use a free API from OpenRouter with a large context window to power the agent behind it.
For a deeper dive into building practical content systems with agents, see our guide on How to Set Up AI Agents for Productivity in 2026: The 6-Step Personal Workflow System.
What Other Free Capabilities Does Hermes Ship With?
Hermes's broader feature set rounds out the picture:
| Feature | What it does | Source |
|---|---|---|
| Persistent memory | MEMORY.md (short-term, ~2,200 char cap), USER.md (user facts, ~1,375 char cap), SQLite FTS5 session search across all past conversations |
Hermes docs |
| Auto-generated skills | The agent writes skills (plain markdown files) as it works and reuses them going forward | Hermes docs — Features Overview |
| Subagent delegation | Spawn up to 3 concurrent subagents (configurable) with isolated contexts; orchestrator-role children can spawn their own workers | Hermes docs — Delegation |
| Multi-platform gateways | Telegram, Discord, Slack, WhatsApp, Signal, Matrix, Teams, email, CLI — start on one, pick up on another | Hermes homepage |
| Cron scheduling | Schedule recurring agent jobs, script-only watchdog pings, and data collection pipelines | Hermes docs — Features Overview |
For guidance on evaluating agents like Hermes on long-horizon tasks, see How to Evaluate AI Agents on Long-Horizon Tasks in 2026.
Hermes Agent Upgrades 2026: Comparison Table
| Upgrade | What it does | Disk/Performance impact | Setup effort | Verified status |
|---|---|---|---|---|
| Session DB optimization | Shrinks state.db by up to 78% (vendor-reported, ~60% average) |
Frees disk; smaller DB = faster reads | Zero (auto if <1GB) | Vendor-reported (source) |
| tldraw offline skill | Agent scripts live whiteboards via local HTTP API | Negligible (a skill file + the desktop app) | Install app + skill | Confirmed (official docs) |
| Kanban multi-agent board | Durable cross-agent task board with swarm topology helper | kanban.db grows per task but is small |
Enable in config | Confirmed (official docs, first-party since v0.15) |
| Hyperframes | HTML-to-MP4 video rendering locally | Output video files; rendering uses CPU | hermes skills install + Node.js/FFmpeg |
Confirmed (official docs) |
| Subagent delegation | Up to 3 concurrent isolated reasoning subagents | Minimal — output is summarized before entering parent context | Built in, no setup | Confirmed (official docs) |
What This Means for You
If you are a developer, solo builder, or small business running AI in your work:
Run
hermes updatetoday. If your database is under 1 GB, the session optimization triggers automatically — you do nothing. If it is larger, runhermes sessions optimize-storageonce after updating. Either way, your search stays live throughout.Try the tldraw skill for visual thinking. When you need to map out an architecture, wireframe a UI, or build a flowchart, ask Hermes to script it on a live tldraw canvas instead of describing it in text. The interactive capabilities (buttons, counters, reactive shapes) mean you can prototype small interactive tools, not just static diagrams.
Use Kanban swarms for multi-step work. Instead of trying to do everything in one long agent session, drop one line on the Kanban board and let specialist agents handle the parts in parallel. The verifier + synthesizer pattern means the result is checked before it reaches you.
Start with one capability, then add. Do not try to set up tldraw, Hyperframes, a Kanban swarm, and multi-platform gateways all on the same day. Pick one — maybe the database optimization (it is already running) or the tldraw skill (it is the most immediately useful) — get comfortable with it, then add the next.
For a broader guide on how AI agents fit into practical workflows, see What to Build With Claude Opus 5 in 2026: The Practical Agentic Playbook and our guide on How to Engineer AI Agent Loops That Are Safe for Production Codebases (2026).
FAQ
Q: How much does the Hermes Agent database optimization reduce disk usage?
A: The optimization reduces on-disk session database size by up to 78% and approximately 60% on average, according to the official Nous Research announcement on July 22, 2026. These figures are vendor-reported and have not been independently verified as of this writing.
Q: Do I need to run a command to trigger the database optimization?
A: Only if your state.db is larger than 1 GB. Databases under 1 GB are optimized automatically on update. For larger databases, run hermes sessions optimize-storage once after updating. The process runs in the background and search remains live throughout.
Q: What is the tldraw offline skill and how is it different from a normal whiteboard?
A: The tldraw offline skill lets Hermes Agent drive the tldraw offline desktop app by writing JavaScript document scripts via a local HTTP API, not by clicking around a GUI. This means the agent can create not just static diagrams but interactive boards — clickable buttons, reactive shapes, working counters — that behave like small apps.
Q: Is the tldraw offline skill included by default in Hermes?
A: No. It is an optional skill. Install it with hermes skills install official/creative/tldraw-offline. You also need the tldraw offline desktop app installed and running with a document open while the agent works.
Q: What is the difference between Hermes Kanban and delegate_task?
A: delegate_task is a function call inside a running agent's session — it does not persist, does not survive restarts, and results go back into the parent's context. Kanban is a durable multi-agent board backed by SQLite where every task is a row, every worker is a full OS process, and work survives crashes and restarts. Use Kanban for work that crosses agent boundaries or needs to outlive a single session.
Q: Can I use a free API with Hermes Agent?
A: Yes. Hermes supports multiple inference providers including OpenRouter, which offers free API access to models with large context windows. You can start with a free OpenRouter key and swap to a paid provider later without changing your Hermes setup.
Discussion
0 comments