vibra-ingenn / live demo real recipes · real timings · captured via mcp
Recipe execution — captured 2026-08-08

Three real recipes ran. Under 200ms each. Here is the output.

This is a demonstration of Vibra-Ingenn — a deterministic local execution engine. The timings and output below were captured by an MCP client calling the running engine. Not a simulation. Not a mockup. The bytes and durations are what actually happened.

What you just saw isn't a chat loop. A traditional agent framework would have re-called the LLM three times, each call re-sending the accumulated conversation, burning tokens and adding round-trip latency.

Vibra takes a different approach. The LLM writes one JSON recipe up front, then steps back. A native Go runtime executes the task graph locally — file operations, HTTP requests, scripts, GUI automation, chained LLM calls — with template variables piping outputs between steps. No context accumulation. No wasted tokens. The recipe is a file on disk you can save, share, version, or re-run.

Below are two more recipes that ran the same way — one that hits a live API, one that queries the filesystem. Same engine, same call signature, real captured output.

Three recipes, side by side

source · MCP capture · adeptuscamini.com
01 File chain — write, read, notify 13mstotal

The simplest chain. Deterministic file I/O with output passed between steps via {{prev_output}}.

recipe.json3 tasks
{
  "workflow_id": "demo_1_file_chain",
  "tasks": [
    {
      "id": "write_greeting",
      "action_type": "file_op",
      "target": "./demo_greeting.txt",
      "payload": {
        "operation": "write",
        "content": "Hello from Vibra..."
      }
    },
    { "id": "read_greeting", ...  },
    { "id": "notify_done", ...  }
  ]
}
captured outputstdout · verbatim
✓write_greetingfile_op0.4ms
✓read_greetingfile_op8ms
✓notify_donenotify1ms
Read returned "Hello from Vibra-Ingenn. This file was written by Claude, executing a real recipe through the vibra-ingenn MCP server..." — 159 bytes, exact round-trip.
02 Live pipeline — fetch, persist, notify 166mstotal

Hits the CoinGecko API, saves the response to disk, then notifies. Real network round-trip captured live.

recipe.json3 tasks · 1 HTTP
{
  "workflow_id": "demo_2_live_btc_pipeline",
  "tasks": [
    {
      "id": "fetch_btc",
      "action_type": "http_request",
      "target": "https://api.coingecko...",
      "payload": { "method": "GET" }
    },
    {
      "id": "save_btc_data",
      "action_type": "file_op",
      "payload": {
        "operation": "write",
        "content": "{{prev_output}}"
      }
    },
    { "id": "notify_saved", ...  }
  ]
}
captured outputlive network fetch
✓fetch_btchttp_request165ms
✓save_btc_datafile_op0.3ms
✓notify_savednotify0.1ms
Response body piped to next step via template: {"bitcoin":{"usd":64921,"usd_24h_change":-0.06}} — persisted to disk, 64 bytes.
03 Filesystem introspection — list, notify <1mstotal

Scans a directory with a type filter, returns the count and paths as JSON. Same call surface as everything else.

recipe.json2 tasks · filter=file
{
  "workflow_id": "demo_3_workspace_scan",
  "tasks": [
    {
      "id": "list_scratchpad",
      "action_type": "file_op",
      "target": "./scratchpad",
      "payload": {
        "operation": "list",
        "type": "file"
      }
    },
    { "id": "notify_result", ...  }
  ]
}
captured outputfilesystem read
✓list_scratchpadfile_op0.1ms
✓notify_resultnotify0.0ms
Returned 15 file paths as a JSON array, with metadata {"count": 15}. Type filter excluded directories.
The idea

The model is your grep.
Vibra is your pipe.

— from the Vibra-Ingenn build journal, day 6

Small local models fail as monolithic all-in-one agents. Treated according to the Unix philosophy — each model call is a specialized primitive, working on a narrow well-defined job — a 3B model operates near 100% accuracy. Vibra is the runtime that chains those specialized calls together, with the same discipline coreutils have had for fifty years: one job per step, output piped to input, everything auditable, everything scriptable.

The recipe is not a chat log. The recipe is a program.

Sits anywhere in your pipeline

tested · closed-source · commercial

Vibra receives from MCP clients and REST clients. It calls out to any OpenAI-compatible LLM endpoint. All three can happen in the same recipe.

Claude Desktop Cursor Windsurf Cline Roo Code Antigravity VS Code MCP REST · localhost:8990 Ollama LM Studio OpenRouter OpenAI Anthropic Gemini Groq Together AI Fireworks vLLM Jan LocalAI browser_op — reserved gui_op — Windows only

Run it yourself.

One-click Windows installer. Auto-registers with Claude Desktop. Works with any local LLM runtime. 7-day free trial, no credit card required to install.

Download installer → View full site
Free tier · keep your work forever Pro$8.99/mo Early adopterrate locked forever