Agent integration docs
This guide explains the roles, capabilities, installation, and practical use of the ElserStudio Skill and local MCP. Once connected, compatible agents such as Codex and Claude Code can read projects, organize Worlds, generate shots, and update the canvas without bypassing the desktop app.
Skill and MCP
The Skill understands the creative goal and organizes the production workflow. MCP executes those steps safely. With only the Skill, an agent knows the method but cannot operate the project; with only MCP, it has tools but not the complete production method.
It organizes novel-to-World, Episode, and Shot workflows with ElserStudio-native modules for shot intent, continuity, references, performance, cinematography, sound, prompting, review, repair, and delivery.
It stores no project state, calls no model directly, and cannot spend by itself.It exposes ElserStudio projects, Worlds, assets, Shots, generation tasks, and canvas capabilities to local agents while sharing the same provider routing and local data as the desktop UI.
Every real read, write, and generation action runs through the active desktop app.Capability scope
The desktop MCP currently provides 60 tools. The official Skill adds native module orchestration, production judgment, reference rules, and spend boundaries on top.
Extract character, location, prop, and voice candidates, then turn a novel or idea into an Episode and editable Shots.
Preserves plot order, dialogue, language, and each shot's narrative job.Create or reuse characters, locations, props, and shared references so identities do not drift between shots.
References bind canonical IDs instead of relying on names in prompt prose.Direct shot size, camera, performance, motion, lighting, style, VFX, sound, and multilingual prompts.
Includes professional vocabulary for English, Chinese, Japanese, Korean, Spanish, and Russian.Use cloud or local BYOK routes for character sheets, scene boards, prop boards, motion guides, and native-audio video.
Submit one task or batches of 2–50 through the same routing as the desktop UI.Poll batches, separate completed, failed, and canceled tasks, and manage shared anchors, sequential shots, and retry variables.
An imperfect result never triggers another paid task without approval.Materialize the story, script, World assets, Shot videos, and Composition as a visible chain while recording Skill modules and quality gates.
ElserStudio's local canonical records remain the source of truth.Current boundary: MCP is available only on the computer running ElserStudio. A client holding the local token has access to the current MCP tool surface, so connect only trusted agents.
Install and connect
The Skill is a public versioned package from this site. Each ElserStudio installation creates its own local MCP token; the website never receives or distributes that credential.
Install and start the latest ElserStudio desktop app.
Configure a cloud account or local BYOK provider in Settings.
Install Node.js 18+ to run npx skills.
A global install lets compatible agents discover elser-studio from any project. The installer reads this site's well-known manifest, downloads the versioned ZIP, and checks its digest.
npx skills add https://elserstudio.ai -g -ynpx skills ls -gnpx skills update elser-studio -g -ySHA-256 07290820a7769d2d4b715e892e9c78ba08c9e89e2f904de42d84e4186a606969
http://127.0.0.1:8787/mcpNever paste the real token into chats, issues, screenshots, or websites. If it is exposed, stop using it and contact support. Keep the desktop app running while using MCP.
Practical use
You do not need to memorize 60 tool names. Tell the agent what to achieve, which Project or World to use, and whether media generation is authorized. The Skill selects the workflow and MCP performs canonical operations.
Provide a novel, chapter, or idea and say whether you need a script, assets, a shot plan, or generated media.
The Skill requires the agent to inspect existing Projects, Worlds, Episodes, Shots, and assets before guessing IDs or creating duplicates.
Text planning may proceed. Images and video need an explicit count, route, and target; open-ended requests default to a small proposed sample.
The agent reports canonical IDs touched, successful and failed tasks, take verdicts, and exactly what needs your review on the canvas.
These requests expand gradually from no-spend planning to controlled generation.
Use the ElserStudio Skill to analyze this novel, reuse the current World, and create an Episode and Shots. Do not generate images or video yet.
Choose three representative Shots from this Episode, verify character, location, and prop references, then generate video with local BYOK and sync it to the canvas.
Keep character identity and dialogue unchanged. Review Shot 4's camera, action timing, lighting, and native sound, then change only the single variable most likely to improve it.
List the recent generation tasks in the current project, separate completed, failed, and canceled results, explain failures, and do not retry paid tasks automatically.
Troubleshooting
Most issues come from a closed desktop app, a session that has not reloaded, mismatched credentials, or provider configuration.
Confirm the desktop app is still running, copy the client configuration again from Settings → Local MCP, and restart the agent session. Do not bypass MCP with direct SQLite access or Provider API calls.
Copy the full connection configuration again. Confirm the header uses Authorization: Bearer with the current token and that the URL remains 127.0.0.1:8787/mcp.
Confirm Node.js 18 or later, then rerun the global install command. You can download the ZIP and compare the SHA-256 on this page; never install modified copies from unknown mirrors.
Ask the agent to query the canonical task and provider task ID through MCP, then check provider balance, model availability, network access, and references. Failed tasks are not retried automatically; confirm the cause and next attempt scope first.
Ready to begin
Start with a short passage, one World, one Episode, and two or three Shots. Complete canonical references and canvas sync before authorizing a few representative generations.