Agent Integration Profiles¶

Connect your favorite coding agent to MoE Sovereign in under five minutes. This page provides quick-start configurations for every supported agent, explains which API endpoint each one uses, and maps out feature compatibility.

How It Works¶

MoE Sovereign exposes two API endpoints:

Endpoint	Protocol	Used by
`/v1/messages`	Anthropic Messages API	Claude Code
`/v1/chat/completions`	OpenAI Chat Completions API	OpenCode, Claw Code, Codex CLI, Aider, Continue.dev, Cursor, Open WebUI

Both endpoints route requests through the same MoE pipeline. The endpoint you use depends solely on what your agent expects — the processing behind it is identical.

Claude Code¶

Claude Code is Anthropic's official CLI agent. It speaks the Anthropic Messages API natively, which means it can use every MoE Sovereign feature without an adapter layer.

Quick Start¶

# ~/.bashrc or ~/.zshrc
export ANTHROPIC_BASE_URL=https://your-moe-instance.example.com
export ANTHROPIC_API_KEY=moe-sk-xxxxxxxx...

# Start Claude Code — all requests now route through MoE Sovereign
claude

After editing your shell profile, run source ~/.bashrc (or source ~/.zshrc) before starting claude.

VS Code Extension¶

If you use the Claude Code VS Code extension instead of the CLI:

{
  "claude-code.apiEndpoint": "https://your-moe-instance.example.com",
  "claude-code.apiKey": "moe-sk-xxxxxxxx..."
}

What Works¶

Capability	Status	Notes
`/v1/messages` endpoint	Fully supported	Native Anthropic protocol
Streaming (SSE)	Fully supported	Real-time token output
Tool use / function calling	Fully supported	File edits, bash, MCP tools
Multi-turn conversations	Fully supported	Full conversation history preserved
Thinking blocks (`<think>`)	Fully supported	Requires `moe_reasoning` or `moe_orchestrated` mode
Image inputs	Supported	Forwarded to vision-capable models
System prompts	Supported	CLAUDE.md, project instructions passed through

Profile Switching via API Key¶

Each API key can be bound to a specific CC Profile in the Admin UI. This controls which MoE mode (native, reasoning, orchestrated) is used for all requests made with that key.

Setup:

Admin UI > Users > select user > API Keys
Set the CC Profile dropdown for the desired key
All requests with that key automatically use the bound profile

This lets you create multiple API keys for different workflows:

# Key bound to cc-ref-native — fast interactive coding
export ANTHROPIC_API_KEY=moe-sk-native-key...

# Key bound to cc-ref-orchestrated — deep research
export ANTHROPIC_API_KEY=moe-sk-deep-key...

Switch between profiles by changing which key is exported.

Limitations¶

Claude Code only speaks the Anthropic /v1/messages protocol — it cannot use the OpenAI-compatible endpoint directly.
Model selection inside Claude Code (/model command) selects from models reported by /v1/models. The actual LLM routing is controlled by the expert template bound to your CC Profile, not by the model name alone.

OpenCode¶

OpenCode is a Go-based, open-source terminal coding agent with support for 75+ LLM providers. It uses the OpenAI-compatible /v1/chat/completions endpoint.

Quick Start¶

Create or edit ~/.config/opencode/config.toml:

[providers.moe]
name = "MoE Sovereign"
base_url = "https://your-moe-instance.example.com/v1"
api_key = "moe-sk-xxxxxxxx..."
models = ["moe-reference-30b-balanced"]

Then start OpenCode:

opencode

Notes¶

OpenCode discovers available models from /v1/models. All expert templates configured in MoE Sovereign appear as selectable models.
Tool use support depends on the OpenCode version. Check the project documentation for the latest capabilities.
Streaming is fully supported.

Claw Code¶

Claw Code is a Python-based, open-source coding agent inspired by Claude Code. It works via the OpenAI-compatible endpoint.

Quick Start¶

export OPENAI_BASE_URL=https://your-moe-instance.example.com/v1
export OPENAI_API_KEY=moe-sk-xxxxxxxx...
export OPENAI_MODEL=moe-reference-30b-balanced

claw-code

Notes¶

Claw Code supports tool use (file edits, bash execution) through the OpenAI function calling protocol.
Streaming is fully supported.
Because Claw Code uses the OpenAI endpoint, CC Profile selection via API key binding is not available. The model name in OPENAI_MODEL determines which expert template is used.

Codex CLI (OpenAI)¶

Codex CLI is OpenAI's official terminal agent. It supports custom base URLs for OpenAI-compatible backends.

Quick Start¶

export OPENAI_BASE_URL=https://your-moe-instance.example.com/v1
export OPENAI_API_KEY=moe-sk-xxxxxxxx...

codex --model moe-reference-30b-balanced

Notes¶

Codex CLI expects a fully OpenAI-compatible backend. MoE Sovereign's /v1/chat/completions endpoint satisfies this requirement.
Tool use and streaming are supported.
The --model flag selects the expert template by its model ID as reported by /v1/models.

Aider¶

Aider is the oldest terminal AI pair-programming tool (39K+ GitHub stars). It supports OpenAI-compatible backends via command-line flags or environment variables.

Quick Start¶

aider --openai-api-base https://your-moe-instance.example.com/v1 \
      --openai-api-key moe-sk-xxxxxxxx... \
      --model openai/moe-reference-30b-balanced

Environment Variables (Alternative)¶

export OPENAI_API_BASE=https://your-moe-instance.example.com/v1
export OPENAI_API_KEY=moe-sk-xxxxxxxx...

aider --model openai/moe-reference-30b-balanced

Notes¶

The openai/ prefix in the model name tells Aider to use the OpenAI provider. This is required.
Aider supports tool use for file editing and git operations.
Streaming is fully supported.
For best results with the MoE pipeline, use a 30b-balanced or larger template — Aider's edit format benefits from strong instruction following.

Continue.dev / Cursor¶

These IDE-integrated agents are covered in the dedicated Continue / Cursor Integration page.

Both use the OpenAI-compatible /v1/chat/completions endpoint and support model selection, streaming, and tool use.

Discovering Available Models¶

All agents can query available models (expert templates) from MoE Sovereign:

curl https://your-moe-instance.example.com/v1/models \
  -H "Authorization: Bearer moe-sk-xxxxxxxx..." | jq '.data[].id'

Each expert template configured in the Admin UI appears as a model in the /v1/models response. Select the model ID that matches your desired quality/speed trade-off.

Compatibility Matrix¶

Feature	Claude Code	OpenCode	Claw Code	Codex CLI	Aider
`/v1/messages` (Anthropic)	Yes	No	No	No	No
`/v1/chat/completions` (OpenAI)	No	Yes	Yes	Yes	Yes
Tool Use	Yes	Limited	Yes	Yes	Yes
Streaming	Yes	Yes	Yes	Yes	Yes
MoE Pipeline	Yes	Yes	Yes	Yes	Yes
CC Profile Selection	Yes (via key)	No	No	No	No
Expert Template Routing	Yes	Yes	Yes	Yes	Yes
Thinking Blocks	Yes	No	No	No	No
Image Inputs	Yes	No	No	No	No

Endpoint determines features

Agents using the Anthropic /v1/messages endpoint (Claude Code) get access to thinking blocks and the full CC Profile system. Agents using the OpenAI /v1/chat/completions endpoint select their expert template via the model name instead.

Troubleshooting¶

"Invalid API Key"¶

curl https://your-moe-instance.example.com/v1/models \
  -H "Authorization: Bearer $ANTHROPIC_API_KEY"

If this returns an error, verify the key in the User Portal under API Keys.

Connection Refused¶

curl https://your-moe-instance.example.com/health

If this fails, the MoE Sovereign instance is not reachable. Check DNS, firewall rules, and whether the service is running.

Model Not Found¶

curl https://your-moe-instance.example.com/v1/models \
  -H "Authorization: Bearer moe-sk-xxxxxxxx..." | jq '.data[].id'

Use one of the returned model IDs in your agent configuration.

Slow Responses¶

Switch to a native CC Profile or a fast expert template (e.g., 8b-fast) for interactive coding.
Use moe_orchestrated mode only for deep research or complex analysis where 2-10 minute latency is acceptable.
See Expert Templates & Profiles for guidance on choosing the right quality/speed trade-off.