openfang-agent-create

OpenFang Agent Create / Configure

Spawn and customize OpenFang agents from natural language. OpenFang's agent registry is in-memory — agents live until the daemon restarts. The 49 templates under vibemind-os/openfang/agents/*/agent.toml.tmpl are the typed starting points; patch + set_tools + set_mcp_servers refine after spawn.

The dashboard at http://127.0.0.1:4200/ (also embedded as the "OpenFang" sub-tab in AgentFarm) reflects every change live. After spawning, the new agent appears in the dashboard's agent list immediately.

Available MCP tools

This skill uses the openfang-agents MCP server (registered in .mcp.json). All 8 tools wrap http://127.0.0.1:4200/api/agents/*:

Tool	When to call
openfang_agents_list	first — see what's already running
openfang_agent_get	inspect one agent's full manifest before patching
openfang_agent_spawn_from_template	create from a named template
openfang_agent_spawn_from_toml	create from a raw TOML manifest (custom config)
openfang_agent_patch	rename / change model / change provider in-place
openfang_agent_set_tools	restrict callable tools (allow/block lists)
openfang_agent_set_mcp_servers	restrict which MCP servers the agent may use
openfang_agent_kill	terminate + unregister

Procedure (5 steps)

1. Look before you leap. Call openfang_agents_list. If an agent with the requested role already exists and is Running, prefer reusing it (return its id to the user). Avoid duplicates.

2. Spawn from template. Prefer openfang_agent_spawn_from_template with one of the canonical names — saves writing a TOML from scratch:

   • brain-coder — code editing, builds, file ops
   • brain-researcher — web search + summarization
   • brain-knowledge — KG / Rowboat / Obsidian queries
   • brain-orchestrator — multi-agent coordination
   • brain-planner — task decomposition
   • brain-devops — docker / git / CI ops
   • brain-security — vuln scans
   • analyst, architect, assistant — general-purpose Capture the returned agent_id.

3. Patch metadata if needed. Display name, description, model and provider can all change via openfang_agent_patch with just the fields you want to set. Skip if the template defaults are fine. System prompt changes require a full respawn (Bewusst nicht in v1).

4. Restrict capabilities if needed. For tight sandboxes:

   • openfang_agent_set_tools with tool_allowlist: ["file_read","web_fetch"] limits the agent to just those tools.
   • openfang_agent_set_mcp_servers with mcp_servers: ["fungus-search"] restricts which MCP servers the agent may connect to. Skip both if the template defaults (often "*" / all) are fine.

5. Verify. Call openfang_agent_get with the new agent_id — confirm state=Running, manifest matches your edits. Then the AgentFarm OpenFang iframe (and the :4200 dashboard) will show the new agent in its list.

Beispielfluss

Erstelle einen Code-Review-Agent basierend auf brain-coder mit fungus-search als MCP-Server.

1. openfang_agents_list
   → [{id:"...", name:"brain-coder", state:"Running"}, ...]
   (already one running — but user wants a NEW one for code-review)

2. openfang_agent_spawn_from_template {template: "brain-coder"}
   → {agent_id: "abc-123", name: "brain-coder"}

3. openfang_agent_patch {agent_id:"abc-123", name:"code-reviewer",
                         description:"Reviews PRs, suggests fixes"}
   → {status: "ok"}

4. openfang_agent_set_mcp_servers {agent_id:"abc-123",
                                   mcp_servers:["fungus-search"]}
   → {status: "ok"}

5. openfang_agent_get {agent_id:"abc-123"}
   → {name:"code-reviewer", state:"Running",
      manifest:{mcp_servers:{servers:["fungus-search"]}, ...}}

Done — der neue Code-Reviewer ist im AgentFarm OpenFang-Tab sichtbar und kann über POST /api/agents/abc-123/message (dashboard chat) angesprochen werden.

Error handling

Every MCP tool returns a TextContent starting with error: on failure — never throws. Typical errors and what to do:

Error prefix	Cause	Action
cannot reach OpenFang at ...	daemon down	tell user to run Vibemind.debug.ps1 -Modules openfang or target/release/openfang.exe start --config openfang.vibemind.toml; do NOT silently retry
OpenFang returned HTTP 404	bad agent_id or template name	re-list first, suggest a similar name to the user
OpenFang returned HTTP 400 ...Provide 'tool_allowlist' and/or...	empty set_tools body	warn user, skip the call (don't infinite-loop)
OpenFang timed out after 60s	LLM call in spawn took too long (unusual)	suggest a smaller template / different model; once it times out, the agent may have spawned anyway — re-list to check

Voice-Pfad

Mit VOICE_BRAIN_MULTIHOP=true in .env routet die Voice-Pipeline „erstelle einen openfang agent für X" durch dieselbe Capability und denselben MCP-Pfad wie Claude Code — keine separate Tool-Registry-Pflege nötig. Der Pfad:

voice → brain (/api/multihop/execute) → cap (openfang_agent_create)
      → MCPExecutor → openfang-agents stdio → :4200/api/agents → Agent läuft

Brain's Phase -2 (BrainMultihopBridge in vibemind-os/voice/python/swarm/routing/) ruft /api/multihop/execute mit dem rohen Intent-Text auf; bei No-Match fällt sie sauber auf Phase -1 (cortex/route → Space) durch.

Bewusst NICHT in dieser Skill-Version

• System-prompt change at runtime — needs PUT /api/agents/{id}/update + full respawn cycle; out of Builder-Fokus scope.
• Session management — handled by the :4200 dashboard / AgentFarm-iframe.
• Sending messages to an agent (POST /api/agents/{id}/message) — that's the user's job via dashboard/iframe or a separate skill (e.g. rowboat_chat equivalent). This skill builds & configures only.

fungus-hook-v6-1779358492