Problem
We want to build application-specific agents using the composable poe-agent runtime: explicitly choose only required plugins and tools, and potentially add our own plugins.
For example, a Slack mention workflow should be able to run an agent with:
- no project/workspace context or filesystem access
- no shell or web plugins
- only a small MCP tool surface (
read_slack_thread, set_official, finalize, etc.)
- optional application-owned plugins later
The underlying package already appears to support the right design via agent().model(...).use(...).mcp(...).run(...) in packages/poe-agent.
Today, external consumers of the published poe-code package cannot use that composition layer directly:
@poe-code/poe-agent is marked private and is not available as a normal dependency.
poe-code does not export the composable builder/plugins from its public package entrypoint.
- The available
poe-agent provider/CLI wrapper selects configuration indirectly and requires an internal cwd, making it a poor fit for application-owned constrained agents.
Request
Expose the composable poe-agent runtime as a supported consumer-facing API, either by:
- publishing
@poe-code/poe-agent, or
- re-exporting a supported composition surface from
poe-code.
The supported API should allow consumers to:
- construct an agent with
agent()
- choose provider/system plugins explicitly
- omit filesystem, shell, web, memory, and spawn plugins entirely
- attach only selected MCP servers/tools
- supply custom consumer-owned plugins
- run without requiring a meaningful repository/project
cwd
Example desired usage
import {
agent,
openaiResponsesPlugin,
systemPromptPlugin,
} from "@poe-code/poe-agent";
const result = await agent()
.model("gpt-5.5")
.use(openaiResponsesPlugin())
.use(systemPromptPlugin())
.mcp({
name: "mention_bot_tools",
command: "bun",
args: ["packages/team-bot-tools/src/mcp.ts", "--role", "official"],
})
.run(prompt);
Why this matters
The composable runtime is substantially safer and easier to reason about for constrained automation agents than invoking a broad coding-agent wrapper and attempting to disable capabilities via project configuration. It also provides the right extension point for application-specific plugins without forking or depending on local source checkouts.
Problem
We want to build application-specific agents using the composable
poe-agentruntime: explicitly choose only required plugins and tools, and potentially add our own plugins.For example, a Slack mention workflow should be able to run an agent with:
read_slack_thread,set_official,finalize, etc.)The underlying package already appears to support the right design via
agent().model(...).use(...).mcp(...).run(...)inpackages/poe-agent.Today, external consumers of the published
poe-codepackage cannot use that composition layer directly:@poe-code/poe-agentis markedprivateand is not available as a normal dependency.poe-codedoes not export the composable builder/plugins from its public package entrypoint.poe-agentprovider/CLI wrapper selects configuration indirectly and requires an internalcwd, making it a poor fit for application-owned constrained agents.Request
Expose the composable
poe-agentruntime as a supported consumer-facing API, either by:@poe-code/poe-agent, orpoe-code.The supported API should allow consumers to:
agent()cwdExample desired usage
Why this matters
The composable runtime is substantially safer and easier to reason about for constrained automation agents than invoking a broad coding-agent wrapper and attempting to disable capabilities via project configuration. It also provides the right extension point for application-specific plugins without forking or depending on local source checkouts.