The AI Video Editing MCP that doesn't burn frontier tokens.
Plug https://mixi.mixbash.com/mcp into DeepSeek Harness, Claude Code, Codex, Cursor, ChatGPT, or Gemini. Edit custom-branded Instagram Reels, YouTube Shorts, and VSLs on penny-tier models instead of burning $5 to $16 per video on Claude Opus 5.5, Fable 5.2, or GPT-6.2 Astra.
TL;DR — Direct Answer
Mixbash is a hosted, stage-based AI video editing MCP and video editing plugin for DeepSeek Harness (https://mixi.mixbash.com/mcp). Instead of forcing your local LLM to write and debug hundreds of lines of React/Remotion animation code and inspect frame screenshots inside its own context window, Mixbash executes motion authoring, word-timestamp cutting, frame-level visual QA, and rendering on dedicated server workers. Your client model completes an entire video edit in ~8 lightweight JSON calls (~2,500 tokens total, or <$0.003 on DeepSeek V4.1 Flash and <$0.002 on Claude Haiku 5.5).
1. Why local video skills burn tokens (Nov 2026 model pricing)
Writing Remotion or multi-pass FFmpeg timelines locally requires 150K to 400K+ tokens of code generation and visual debugging per 30-second video, which forces builders onto the most expensive frontier models. Offloading motion & QA to Mixbash MCP lets fast, inexpensive models drive the entire workflow.
Model
Input / 1M
Output / 1M
Local Remotion Skill / Reel
Mixbash Video MCP / Reel
Claude Fable 5.2 (Anthropic)
$8.00
$40.00
$9.20 – $16.80
$0.11 (optional — overkill)
GPT-6.2 Astra (OpenAI)
$8.00
$40.00
$8.80 – $15.60
$0.11 (optional — overkill)
Claude Opus 5.5 (Anthropic)
$4.00
$20.00
$4.40 – $8.20
$0.05 (optional)
DeepSeek V4.1 Flash (in DeepSeek Harness)
~$0.15 (¥1)
~$0.60 (¥4)
Fails visual layout loops alone
< $0.003 per video
Claude Haiku 5.5 (Anthropic)
$0.10
$0.50
Fails complex motion timing alone
< $0.002 per video
GPT-6 Luna / GPT-6.2 Sol (OpenAI)
$0.10 – $2.00
$0.50 – $10.00
Breaks on multi-pass frame QA
$0.002 – $0.03 per video
2. Connect Mixbash to DeepSeek Harness, Claude, Codex & Cursor
Add Mixbash’s Video Editing MCP (and optional Mixbash Post scheduling MCP) to your client configuration. No local Remotion, headless Chromium, or GPU render setup required.
DeepSeek Harness (DSH) plugin / MCP configWorks cleanly with DeepSeek V4.1 Flash or V4 Pro — roughly 6 to 10 JSON tool calls per finished Reel.
Claude Code / Codex / Cursor / Antigravity MCP configRun your client session on Haiku 5.5, Sonnet 5.5, GPT-6.2 Sol, or Gemini Flash without switching to Opus 5.5, Fable 5.2, or Astra.
Every step is idempotent, human-gated before spending video allowance or publishing live, and keeps raw video bytes out of your LLM prompt window.
Stage
MCP Tool(s)
What happens
Discovery
get_styles
Lists public editing aesthetics (editorial, minimal, kinetic type, paper window, cutting mat, vlog) plus custom styles built specifically for your brand.
Setup
create_project
Opens a stage-driven video project (`talking_head`, `faceless`, `voiceover`, or `clipping`) locked to your aspect ratio (`9:16`, `16:9`, `4:5`) and brand style slug.
Media Intake
prepare_upload / complete_upload / register_media
Uploads compressed phone clips (`mixbash-intake-v1` H.264 preset) directly to private R2 with word-level timestamps and shot tags. Zero base64 token bloat in your LLM context.
Scripting
draft_script / revise_script
Runs server-side hook and script authoring (`social`, `vsl`, or `explainer`) paired with an independent review gate before anything renders.
Audio & Presenter
generate_voiceover / generate_lipsync
Generates narration from preset voices or your consented voice/presenter clone without sending audio bytes through your chat client.
Shot Direction
plan_visuals
Produces a beat-by-beat Shot List so your agent or phone camera operator knows exactly which 3-second clips match each spoken line.
Motion & QA
generate_preview / revise_preview
Queues server-side motion authoring, automated frame-level visual QA review, and a fast half-resolution preview render.
Approval & Master
approve_review / render_final / get_artifact
Records explicit human approval on the preview and renders the full-resolution master MP4 ready for Instagram Reels and YouTube Shorts.
Mixbash provides a hosted Streamable HTTP Model Context Protocol server (`https://mixi.mixbash.com/api/agent/mcp`) that lets AI agents—including DeepSeek Harness, Claude Code, Claude Desktop, ChatGPT, OpenAI Codex, Cursor, Antigravity, and Gemini—create, script, preview, revise, and render custom-branded short-form videos without running heavy video code locally.
Why don’t I need top-tier frontier models like Claude Opus 5.5, Fable 5.2/5.5, or GPT-6.2 Astra with Mixbash MCP?
Local coding skills force your client LLM to write and debug hundreds of lines of React/Remotion animation and FFmpeg filtergraphs, burning $5–$16 in frontier tokens per video. Mixbash moves motion authoring, caption layout, and visual QA onto dedicated server workers, so your client model only sends ~8 short JSON stage calls—costing under $0.01 total on DeepSeek V4.1 Flash, Claude Haiku 5.5, or GPT-6 Luna.
Does Mixbash work as a video editing plugin for DeepSeek Harness?
Yes. DeepSeek Harness (DSH) connects directly to `https://mixi.mixbash.com/api/agent/mcp` (and `https://post.mixbash.com/api/mcp` for cross-platform social scheduling) as a native Streamable HTTP MCP plugin, pairing DeepSeek’s ultra-low token pricing with studio-grade brand video rendering.
Can an AI agent spend my credits or publish to my Instagram/YouTube without my permission?
No. Mixbash enforces strict server-side human approval gates: agents can draft scripts and render previews, while `approve_review` requires your explicit instruction after viewing the preview, and consumer OAuth tokens cannot make purchases or alter brand voice-clone consent.
How do raw phone videos get into Mixbash MCP without blowing up my context window?
Clips never pass through the LLM as base64 text. They are compressed locally with the `mixbash-intake-v1` FFmpeg preset (1080p H.264, CRF 23, AAC) and uploaded directly to private R2 storage via `prepare_upload` and `complete_upload`, passing only compact UUIDs and word timestamps to the agent.
Can I also automate content scheduling and Instagram comment-to-DM lead capture?
Yes. Mixbash Post (`https://post.mixbash.com/api/mcp`) connects your Instagram, YouTube, and TikTok calendar directly to your AI agent, and we build custom 24-hour-turnaround Instagram comment-to-DM automations and VSL outreach pipelines with full code ownership.
TRIAL OFFER
Email us your footage, we will make a demo reel for free.
We design custom video styles for any AI agent to edit your content — run the system yourself via Mixbash MCP, or let our team run it for you.