Eden AI/ research
Capability compositionv0.1

Give any model access to any capability.

a text modelmcpevery specialist toolresult

Scroll to gather
phase 01Isolation

A model, alone.

A text model is superb at text. Images, audio, scanned pages and the live web sit outside its native reach.

native
text
reach
0 / 8
phase 02Capability gap

The capabilities exist. Just not here.

OCR, vision, speech, web access: each is a solved problem, running somewhere else. The model cannot touch any of it.

specialists
8
distance
∞
phase 03MCP

A routing layer between them.

Eden AI’s MCP server sits between the model and every specialist. One protocol, many tools. The model only needs to know how to ask.

protocol
mcp
tools exposed
8
phase 04Composition

One call. Out and back.

The model requests a capability. The request travels through MCP to a specialist; the result travels back as text the model already understands.

path
model → mcp → tool → model
return
result
phase 05Accumulation

Every capability feeds the core.

Each tool the model can call adds energy. A cheaper, smaller model starts doing work that used to require something far larger.

streams
8 / 8
reach
8 / 8
phase 06Genkidama

One system. Many sources.

Gathered from many specialists, the composite is more capable than any single model in it. That gathering is the whole idea.

state
composite
energy
max
01The ideaeden ai · research

A model does not need to possess every capability.
It needs access to it.

Genkidama is an Eden AI research project: a text-oriented model reasoning over what specialist tools give it, from OCR and image analysis to transcription and web research, through Eden AI's MCP server.

The name is the Spirit Bomb: many sources of energy, gathered into one.

02The gap8 capabilities · 3 configurations

Where the reach ends.

A capable model can still be unable to inspect an image, transcribe audio, read a scanned page or search the web on its own. The limit is not reasoning; it is direct access. One catalogue, three positions: a text-only model, a natively multimodal model, and the text-only model with Eden AI's MCP server in reach.

native
out of reach
via mcp specialist
capabilitytext-only modelmultimodal modeltext-only + mcp
OCOCReden.ocr.extract
IMImageeden.image.describe
AUAudioeden.audio.analyze
SPSpeecheden.speech.transcribe
VDVideoeden.video.analyze
WBWebeden.web.fetch
TRTranslationeden.text.translate
DCDocumentseden.document.parse
03The bridgemodel context protocol
architecturemodel · mcp · 8 families · 39 tools
Genkidama architectureA text-only model connects to Eden AI's MCP gateway, which routes tool calls to 39 expert-model tools grouped in 8 capability families, and returns results to the model.MODELtext-onlyMCPeden ai gatewayIMVISION11 toolsOCOCR & DOCUMENTS6 toolsAUAUDIO2 toolsVDVIDEO2 toolsWBWEB7 toolsTXTEXT6 toolsTRTRANSLATION2 toolsUTUTILITY3 toolstool_callresult
Vision11
object detectionface detectionface recognitionface comparelogo detectionexplicit contentAI-image detectiondeepfake detectionimage generationbackground removalanonymization
OCR & documents6
OCROCR, asynctable extractionfinancial parseridentity parserresume parser
Audio2
speech-to-texttext-to-speech
Video2
deepfake detectionvideo generation
Web7
searchresearchscrapingbatch scrapecrawlsite mapstructured extraction
Text6
moderationAI-text detectionentity recognitiontopic extractionspell checkplagiarism detection
Translation2
automatic translationdocument translation
Utility3
upload filecheck joblist models
catalogue discovered at runtime via tools/listsnapshot 2026-09

One protocol between the model and everything it lacks.

MCP is a standard interface between a model and external tools. Instead of building every capability into the model, an agent discovers the tools a server exposes, calls the one it needs and reads the result back as context.

Eden AI's hosted MCP server exposes its expert models this way: one tool per feature, plus a few utilities. OCR, image analysis, web search, speech, translation and the rest of the catalogue become callable by any MCP-capable agent, and the catalogue is discovered at runtime rather than hardcoded.

  • discoverytools/list
  • invocationtools/call
  • returntext · structured
Read the MCP server documentation

One call, out and back.

The model decides it needs a capability and emits a tool call. The gateway routes it to the specialist; the result returns as text, and the model keeps reasoning with it. One example, through OCR:

tool-call trace
  1. MODELtext-only · decides it needs OCR
  2. MCPeden ai gateway · routes the call
  3. OCRocr · reads the image
  4. RESULTtext · returned as context
  5. MODELcontinues reasoning
tool_call.jsonrequest · example
{
  "method": "tools/call",
  "params": {
    "name": "ocr",
    "arguments": {
      "file": "receipt-014.jpg",
      "language": "auto"
    }
  }
}
example session · illustrative5 events
  1. 00.000prompt "What is the total on this receipt?"
  2. 00.412model text-only · cannot read the image
  3. 00.418tools/call ocr → mcp gateway
  4. 00.830ocr text returned
  5. 01.204answer "The total is €48.20."
04The next problemwhat the experiments investigate

If capability can be gathered,
what should a model be?

  1. 01Which capability does the task actually need, and when is a call worth its cost?
  2. 02How many tools to call, and how to combine what they return?
  3. 03When to stop, answer, or recover from a tool that failed?
Eden AI ↗MCP serverBack to the core