plugins and MCP

a plugin is a folder with skills and MCP servers. bise loads it in every agent, local stdio servers and remote ones, with their logins.

bise loads Agent Plugins 1.0: a folder with a plugin.json, an optional skills/ folder and an optional mcp.json. the same plugin works in every agent.

#where plugins live

a plugin folder:

my-plugin/
  plugin.json
  skills/
    review/SKILL.md
  mcp.json

{
  "$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json",
  "name": "my-plugin",
  "version": "0.1.0",
  "description": "what it gives the agents"
}

its skills show as my-plugin:review, and its MCP tools as tools.my_plugin.<tool>.

#see and change them

bise plugins                    # every plugin, its state, and what's wrong with it
bise plugins disable my-plugin
bise plugins enable my-plugin

in bise, /plugins does the same. the open agents pick up a change at their next idle.

#MCP servers

mcp.json lists the servers. it takes the same shape as Claude Code's, Cursor's and Vibe's .mcp.json, so a file copied from them works as it is.

#local servers (stdio)

{
  "mcpServers": {
    "files": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "${PLUGIN_ROOT}/data"],
      "env": { "LOG_LEVEL": "warn" }
    }
  }
}

• command is a program on your PATH, or ./x inside the plugin.
• ${PLUGIN_ROOT} is the plugin's folder; ${PLUGIN_DATA} a folder of its own that survives updates, ~/.bend-harness/plugin-data/<name>/.
• a server gets your environment, plus PLUGIN_ROOT, PLUGIN_DATA and its env.

#remote servers

{
  "mcpServers": {
    "linear": { "type": "http", "url": "https://mcp.linear.app/mcp" },
    "internal": {
      "type": "http",
      "url": "https://mcp.example.com/mcp",
      "headers": { "Authorization": "Bearer ${INTERNAL_MCP_TOKEN}" }
    },
    "legacy": { "type": "sse", "url": "https://example.com/sse" }
  }
}

http (or streamable-http) is Streamable HTTP; sse the older HTTP+SSE. ${VAR} and ${VAR:-default} come from your environment when the server connects. a variable that isn't set leaves that server out, and bise plugins says which.

#logging in (OAuth)

a remote server with no Authorization header can log you in with your browser:

bise plugins login            # the servers that can log in
bise plugins login linear
bise plugins logout linear    # forget its tokens

in bise: /plugins login. the server's tools reach your agents as soon as you're logged in, no restart. the tokens stay in ~/.bise/secrets/mcp-oauth/, readable only by you; they are never printed or sent to a model.

with bise on another machine (SSH), log in in your local browser. the last page won't load: copy its address, then paste it into bise plugins login, or run /plugins login linear <address>.

for a server with no dynamic registration (GitHub, Slack), give the client in mcp.json:

{
  "mcpServers": {
    "github": {
      "type": "http",
      "url": "https://api.githubcopilot.com/mcp/",
      "oauth": { "clientId": "your-client-id", "callbackPort": 8765 }
    }
  }
}

#server options

these keys work on any server, local or remote (Codex's names):

#bring your servers from Claude Code or Codex

bise plugins import-mcp my-servers --dry-run < servers.json
bise plugins import-mcp my-servers < servers.json

it turns a list of MCP servers (Claude Code's or Codex's) into one plugin named my-servers. it keeps remote servers, their headers and their login clients, and prints host and header names, never a value.

#how agents use the tools

MCP tools are not all put in the model's prompt. an agent searches for the tool it needs, reads its description, and calls it from a small TypeScript program. so many servers cost few tokens until they are used.

#not supported yet

hooks, agents, knowledge, views, connectors.json and libraries.json of the spec are listed by bise plugins as not supported; the rest of the plugin still loads.