OpenFusion

Showcase

OpenFusion + Seepient ​

Seepient (seepient.zyntopia.com) is our own agent — a lightweight, local-first AI agent CLI from the same author as OpenFusion. It's the agent we use to showcase OpenFusion, because the division of labor between the two is exactly what fusion is built for: Seepient does the legwork; OpenFusion brings the panel.

This page covers setup (about five minutes if OpenFusion is already configured) and a complete use-case walkthrough.

Why this pairing ​

  • Same beliefs. Seepient runs on your machine, uses your provider keys, and asks before acting — its consent modes range from ask-everything to fully autonomous.
  • Native MCP hosting. Seepient's gateway connects MCP servers as first-class tool sources. OpenFusion plugs in without hacks or wrappers.
  • Same skill format. Seepient reads SKILL.md + references/ folders, so the OpenFusion agent skill works unmodified.

1 · Install Seepient ​

bash
brew tap hashangit/seepient
brew install seepient
seepient setup        # one-time wizard: provider keys, model assignments

Full installation and configuration docs live at seepient.zyntopia.com.

2 · Register OpenFusion as a gateway target ​

Seepient's gateway is enabled by default and reads its targets from ~/.seepient/gateway.json. Add OpenFusion to the targets array:

json
{
  "targets": [
    {
      "kind": "mcp",
      "transport": "stdio",
      "command": "npx",
      "args": ["-y", "openfusion-mcp"],
      "description": "Fusion panel — fans a prompt out to several models and returns one judged answer",
      "tags": ["ai", "research", "review"],
      "enabled": true
    }
  ]
}

Then restart Seepient (or start a new session) so the gateway picks it up. If you've disabled the gateway in settings, re-enable gateway.enabled and restart.

Building OpenFusion from source?

Point the target at your build instead — "command": "node" with "args": ["/absolute/path/to/OpenFusion/dist/index.js"].

Prefer the REPL? The same registration works from inside a Seepient session:

/gateway add openfusion {"kind":"mcp","transport":"stdio","command":"npx","args":["-y","openfusion-mcp"],"description":"Fusion panel","tags":["ai"],"enabled":true}

Check that it landed:

/gateway list

You should see openfusion (enabled, MCP/stdio) with the discovered tools counted. /gateway toggle openfusion disables it without removing it; /gateway remove openfusion unregisters it. Seepient exposes the connection through its gateway tools (gateway_call_tool, gateway_route, …) — the agent discovers and calls fusion through them, with semantic routing picking fusion for questions that deserve a panel.

3 · Install the agent skill (recommended) ​

Seepient speaks the same skill format OpenFusion ships. Install it globally so every workspace benefits:

bash
mkdir -p ~/.seepient/skills/openfusion
cp -r "$(npm root -g)/openfusion-mcp/skill/"* ~/.seepient/skills/openfusion/

For a single project, copy into the project's .agents/skills/openfusion/ instead. The skill teaches Seepient the working method — gather context first, then call fusion once with a prepared dossier.

4 · Verify ​

Start Seepient and hand it a fusion-worthy question:

text
seepient
> Research how this worker processes jobs, then ask the fusion panel whether a
  queue or a streaming pull is the better fit. Include the code you found.

Seepient does the legwork, routes the prepared question to OpenFusion through the gateway, and you'll see progress as each candidate answers and the judge synthesizes. The run lands in the OpenFusion dashboard at http://localhost:9077 like any other.

Showcase: an architecture decision, end to end ​

This is the demo we run when we show the pairing. Scenario: which cache invalidation strategy should this service use?

  1. The legwork — Seepient. Ask it to research before answering. It reads the service, enumerates the read/write patterns, notes the consistency requirements, and drafts a dossier. All local, all with its own model — no fusion cost yet.

  2. The panel — OpenFusion. Seepient calls fusion once, with the dossier as context. OpenFusion fans the prepared question out to your configured panel in parallel, drops candidates that stall, and the judge compares the survivors — consensus, contradictions, blind spots — before writing one answer.

  3. The receipt. Open the dashboard: the fusion is logged with a row per sub-call, so you can see which model said what, what it cost, and how long it took.

The point the demo makes: the agent is responsible for what to ask; the panel is responsible for what the answer should be. Neither duplicates the other.

Next up