Providers

Providers

Multi-provider support for Claude Code, OpenAI Codex CLI, and Google Gemini CLI.

---

Overview

Loki Mode supports three AI providers with different capability levels:

Provider	Status	Task Tool	Parallel	MCP	Context
Claude	Full	Yes	Yes (10+)	Yes	200K
Codex	Degraded	No	No	No	128K
Gemini	Degraded	No	No	No	1M

---

Claude Code (Default)

Full-featured provider with complete Loki Mode capabilities.

Installation

# Install Claude Code CLI
npm install -g @anthropic-ai/claude-code

# Authenticate
claude login

Models

Tier	Model	Use Case
Planning	claude-opus-4-7	Architecture, system design (1M context, adaptive thinking)
Development	claude-sonnet-4-6	Implementation, testing (1M context)
Fast	claude-haiku-4-5	Simple tasks, monitoring (200K context)

Invocation

# Launch Claude with autonomous permissions
claude --dangerously-skip-permissions

# In Claude:
# "Loki Mode with PRD at ./my-prd.md"

Capabilities

• Task Tool - Spawn subagents for parallel work
• Parallel Agents - Up to 10+ concurrent agents
• MCP Integration - Extended tool capabilities
• Extended Thinking - Deep reasoning for complex problems
• 3 Model Tiers - Right-size for each task

Configuration

# Set as default provider
loki provider set claude

# Or via environment
export LOKI_PROVIDER=claude

---

OpenAI Codex CLI

Degraded mode with sequential execution only.

Installation

# Install Codex CLI
npm install -g @openai/codex-cli

# Authenticate
codex auth

Model

Model	Context	Notes
gpt-5.3-codex	128K	Official model for Codex CLI v0.98+

Invocation

# Recommended (v0.98.0+)
codex --full-auto

# Legacy
codex exec --dangerously-bypass-approvals-and-sandbox

Limitations

• No Task tool (sequential only)
• No parallel agents
• No MCP integration
• Single model (uses effort parameter)

Configuration

# Set as provider
loki provider set codex

# Or via environment
export LOKI_PROVIDER=codex

# Start with Codex
loki start ./prd.md --provider codex

Effort Parameter

Codex uses an effort parameter instead of model tiers:

effort: low    -> Quick responses
effort: medium -> Balanced (default)
effort: high   -> Thorough analysis

---

Google Gemini CLI

Degraded mode with large context window.

Installation

# Install Gemini CLI
npm install -g @google/gemini-cli

# Authenticate
gemini auth

Model

Model	Context	Notes
gemini-3-pro-medium	1M	Placeholder name

Invocation

# Autonomous mode (verified v0.27.3)
gemini --approval-mode=yolo "Your prompt here"

# Note: -p flag is DEPRECATED - use positional prompts instead

Limitations

• No Task tool (sequential only)
• No parallel agents
• No MCP integration
• Single model

Configuration

# Set as provider
loki provider set gemini

# Or via environment
export LOKI_PROVIDER=gemini

# Start with Gemini
loki start ./prd.md --provider gemini

---

Provider Management

Check Current Provider

loki provider show
# Output: Current provider: claude

List Available Providers

loki provider list
# Output:
# Available providers:
#   claude  (installed, default)
#   codex   (installed)
#   gemini  (installed)

Get Provider Info

loki provider info claude
# Output:
# Provider: claude
# Status: Full features
# Model: claude-opus-4-7
# Context: 1M tokens
# Capabilities: Task tool, parallel, MCP, adaptive thinking

Set Default Provider

# Persists across sessions
loki provider set codex

Per-Session Override

# Override for single session
loki start ./prd.md --provider gemini

---

Feature Comparison

Task Tool (Subagents)

Claude: Full support

Spawn up to 10+ parallel subagents for:
- Research tasks
- Code review
- Testing
- Documentation

Codex/Gemini: Not supported

All tasks run sequentially in main context

Parallel Execution

Claude: Git worktrees + parallel agents

export LOKI_PARALLEL_MODE=true
export LOKI_MAX_PARALLEL_SESSIONS=3

Codex/Gemini: Sequential only

Each task completes before next begins

Context Window

Provider	Context	Effective Use
Claude	200K	Large codebases
Codex	128K	Medium projects
Gemini	1M	Very large files

---

Degraded Mode Behavior

When using Codex or Gemini:

1. No Parallel Agents - Tasks run sequentially
2. No Task Tool - Cannot spawn subagents
3. No MCP - Limited to built-in tools
4. Single Model - No tier selection
5. Longer Execution - Same work takes more time

Automatic Fallbacks

Loki Mode automatically adjusts when in degraded mode:

• Phases run sequentially instead of parallel
• Code review uses single pass instead of 3-reviewer
• Research tasks inline instead of background

---

Provider Selection Guide

Use Claude When:

• Complex multi-file changes
• Need parallel execution
• Require code review quality
• Using MCP integrations
• Speed is important

Use Codex When:

• OpenAI ecosystem preference
• Simpler, focused tasks
• Cost optimization needed
• Sequential workflow acceptable

Use Gemini When:

• Very large context needed
• Google ecosystem preference
• Simple tasks with large files
• Cost optimization needed

---

Troubleshooting

Provider Not Found

loki provider info codex
# Error: Provider 'codex' not installed

# Solution: Install the CLI
npm install -g @openai/codex-cli

Authentication Failed

# Re-authenticate
claude login
codex auth
gemini auth

Wrong Provider Used

# Check current provider
loki provider show

# Reset to default
loki provider set claude