Fully Automated HeyGen Avatar Editing: A Q&A Guide with Claude and Agent OS

Why did fully automatic avatar-video editing only become practical in the last week? In this 27 September Q&A the narrator walks through how raw HeyGen avatar clips can now go from flat talking head to publish-ready cut with no manual editing. Raw output is a single locked take with no cuts, rhythm or emphasis. The fix is a desktop orchestrator that handles both generation and edit in one loop, and the narrator tests it live.

First, what raw HeyGen actually is. HeyGen turns text into a photorealistic avatar speaking, but the file arrives as one continuous shot. The narrator opens a 10-minute example to show there are no trims or colour tweaks. At this point the HeyGen API matters: according to the MCP documentation on developers.heygen.com, HeyGen exposes a Model Context Protocol server that authenticates with an API key and offers tools to create a video, poll status and download the result. The raw clip is therefore not just a media file but a programmable object you can drive from Claude.

Automated editing with Claude: two practical paths

The narrator outlines two fully automated paths. One, hand an already generated HeyGen clip to Claude Desktop running an Opus-class model and say edit this like usual. Two, generate from scratch: give Claude the HeyGen MCP key and docs and ask in one prompt to both create and edit. The example prompt, as described on the Claude integration page on heygen.com, translates naturally: create a 30-second HeyGen avatar video about Z code and edit it in the usual style. Claude first calls HeyGen to render, then applies its own edit skills to cut it. The same loop works for landscape or portrait.

Timings are honest and important for planning. A 30-second clip takes about 10 minutes to edit with Opus, a 10-minute video needs 30 to 60 minutes. The trade-off is clear: re-motion and design-editing skills need time to generate transitions, caption placement and emphasis shots frame by frame. The narrator notes this stack only stabilized about a week ago; earlier attempts often broke at the MCP connection or in the edit chain.

The API-key flow is practical and cautions about secrets. Steps are: go to developers.heygen.com, create an API key, grant all permissions, set an expiry, name it Opus 5.5. Then in Claude Desktop fork the conversation and send HeyGen API key equals. The narrator recommends putting the key in an ENV file rather than pasting it in chat, then attaching the API docs so Claude loads them into context. After that Claude can call HeyGen create and edit tools on its own and deliver a final cut that in the demo looks clean and rhythmic even with audio muted, produced in 5 to 10 minutes for a short clip.

A verification note is needed here. The names Opus 5.5 and Opus 515 repeated in the video do not exist as shipped products. According to the Claude 4 announcement on anthropic.com, the current family is Claude 4, with Opus 4 and Sonnet 4 introduced on 22 May 2025 as the most capable models supporting a 200k context window. The workflow described is valid, but the model name should be read as future-leaning or codename; the key point is that edit quality is only stable on flagship-scale models and will not replicate on every small model.

Agent OS, skills and multi-profile Hermes

The second block answers Andrea on updating Agent OS. The advice is to grab the latest zip from the Agent OS section inside the classroom, return to the environment where you originally installed it with Claude or Codex, start a new chat, select the install folder and attach the zip. The prompt is simple: can you update Agent OS using the update file. The system works in the background overwriting files and then you initialize and test. This avoids a full reinstall and keeps existing data in place.

When to make a skill is the most useful part. The rule is any work you do daily deserves a skill, exemplified by SEO work. The narrator gives a three-step method: first describe what you want the agent to do, after a successful run say save this workflow as a skill so we can reuse it, then test and iterate with feedback such as update the skill file based on this comment. Saved as an MD file you can port the skill between Codex and Claude. The caution is crisp: do not skill everything or the agent bloats and scatters; turn only daily flows into persisted skill architecture , leave one-offs as prompts.

Louis reports a concrete Hermes gateway failure on a Mac mini: cannot switch between OpenAI and Gemini in the gateway, free quote code fails, AI does not start. The fix is not to reinstall but to ask the agent that set up Agent OS to repair. An example prompt asks to ensure switching between OpenAI and Gemini inside the Hermes gateway, set up three distinct profiles inside Hermes, test they actually work and patch bugs, and ask for any API keys needed. According to the Profiles guide on hermes-agent.nousresearch.com, Hermes can run multiple independent agents on the same machine, each with its own config, keys, memory, sessions, skills and gateway state, stored under ~/.hermes/profiles/ . The live demo shows the OpenAI profile initially failing then passing after login and verification, reinforcing the value of incremental communication rather than requesting everything at once.

Jeremy asks for a hybrid setup across several computers: local profiles stay per-device, a cloud profile is shared, Obsidian handles memory backup. The answer is yes, Hermes supports it. You can tell Claude Desktop to add a new profile for this model inside my Hermes agent, here is the API key or local model details, and the system wires it and adds it to the dropdown. To link several machines the narrator prefers a Tailscale mesh over a raw VPS; a 10-minute tutorial gives all agents remote access to the same Agent-X OS, even from a phone. Obsidian sync is orthogonal to profiles; with an account it syncs everywhere as a separate memory layer, which is the correct mental model rather than a model key.

The final section is the most technical: Almir reports five local models failing in five different ways under Hermes on an RTX 5070 and asks about thinking mode on 3.5 4B, AOP yarn context extension, VLM versus llama.cpp runtime, grammar versus schema output constraints, two-stage architecture, and which 8-14B model with 16 GB and 64k context has the best multi-turn tool-use record. The narrator deliberately stays non-technical, recommends letting Claude Desktop take control of the machine and iterate until tests pass, and says running with Ollama defaults is usually enough. For a small model the pick is LFM 2.5 2.6B. According to the technical page on docs.liquid.ai, this model uses a hybrid architecture at 2.6 billion parameters, is optimized for edge devices and delivers fast inference. According to the model card on huggingface.co, LiquidAI/LFM2.5-2.6B is particularly strong at tool calling and because the Hermes framework was used in training data it runs quickly in agent harnesses; it was tested on Mac Studio and found smooth. For the 8-14B bracket the narrator admits no RTX 5090 testing and sticks to personal experience.