Allye Plugin

A methodology and context layer for supported AI coding agents.

The Allye Plugin brings Allye's workflow guidance, skills, memory, and delivery practices into supported AI coding agents. It does not replace the agent: it gives the agent a structured way to understand intent, load the right guidance, act with appropriate checkpoints, and verify the result.

The plugin is distributed from the allye-app/allye-plugin repository. Its integrations are adapted to each host so that Allye works with the host's own strengths and conventions.

Which platform should I choose?

Pi is the recommended platform for using the Allye Plugin. Pi exposes an adaptive toolkit that lets Allye and the host choose the smallest useful workflow, while detecting optional capabilities such as delegation and external runtimes. It is the best starting point when you want the harness to be configured around Allye's workflow rather than follow a fixed host pattern.

The plugin is tested extensively with Claude Code, Codex, and OpenCode. All three are supported choices, with different integration models and different levels of host control:

  • Pi and OpenCode give the agent more autonomy to choose and configure how the harness works. OpenCode also provides a native set of selectable Allye agents.
  • Claude Code offers a deep native integration, including hooks and bounded subagents. Because Claude Code exercises more control over when context and tools are used, you may sometimes need to explicitly ask it to access Allye Memory or load a particular Allye skill. This is a host behavior, not a limitation of the workflow, and Claude Code remains a strong choice when its native structure fits your team.
  • Codex uses Allye through MCP and an instruction manifest. It keeps one host agent and is a straightforward choice when you prefer Codex's conversational model.

Other supported hosts use the same capabilities through their own integration style:

PlatformIntegration modelWhat is installedAgent shape
Claude CodeNative pluginMarketplace plugin, session hook, MCP configuration, skills, and five Claude subagentsInteractive workflow skills plus five bounded subagents
OpenCodeNative package plus MCPallye-opencode package, MCP configuration, and six picker agentsSix selectable Allye agents
CodexMCP plus manifestMCP configuration and AGENTS.mdOne host agent with Allye instructions
CursorMCP plus manifestMCP configuration and .cursorrulesOne host agent with Allye instructions
Gemini CLIMCP plus manifestMCP configuration and GEMINI.mdOne host agent with Allye instructions
Hermes AgentAdapter plus disk skillsMCP configuration, exported skills, and a bootstrap pluginOne host agent with disk-based skills
PiNative packageallye-pi package and canonical skill resourcesAdaptive toolkit with capability detection

The same workflow principles apply across platforms, but the distribution mechanism and available host features differ. Native delegation, a picker, or a bootstrap hook is an optimization—not a prerequisite for planning, implementation, review, handover, or verification.

What the plugin provides

Adaptive workflow guidance

Allye is a toolkit, not a mandatory phase chain. The agent chooses the smallest useful loop for the request:

intent → context → optional research → decision/consent → action → verification → optional persistence

Checkpoints can be skipped or repeated proportionally. Meaningful, delegated, shared, or review-heavy work is encouraged to use Allye work items for traceability. A small local change can use an explicitly approved no-task path.

Skills loaded when needed

Skills are reusable playbooks for discovery, planning, implementation, review, delivery, handovers, and verification. They are loaded on demand instead of forcing every conversation through every phase. See Skills for the complete inventory and availability by platform.

Context and continuity

When the Allye connection is available, the agent can load relevant team context, discover team skills, search prior decisions, and save durable session state. If an optional capability is unavailable, the integration should say so and continue locally when that is safe; it must not pretend that context or verification happened.

Verification and review

The workflow keeps implementation and completion claims evidence-based. Deterministic behavior uses the Red → Green → Refactor discipline when appropriate, and the bounded verification loop repeats test, fix, and test until the result is understood. Native integrations may provide additional delegation or review roles, but local execution remains valid.

Remote MCP connection

The Allye MCP connection is provided remotely through a URL and browser-based OAuth. Add the MCP URL supplied by the host integration or administrator, complete the sign-in and consent flow in your browser, and let the host manage the resulting session. You do not need to paste credentials into a prompt or configuration file.

The connection model is the same across supported platforms, while each host presents it differently. See Installation for host-specific steps and Workflow for guidance on choosing how much control to delegate to the host.

Next steps