Configuration Guide

VibeCody uses TOML-based configuration for VibeCLI. VibeUI provider settings are managed through environment variables or the in-app settings UI.

VibeCLI Configuration

Location: ~/.vibecli/config.toml

The file is created automatically with defaults on first run. You can also create it manually.

Full Reference

# ── Providers ──────────────────────────────────────────────────────

[ollama]
enabled = true
api_url = "http://localhost:11434"   # Local Ollama endpoint
model = "qwen3-coder:480b-cloud"     # Any model pulled via 'ollama pull'

[claude]
enabled = false
api_key = "sk-ant-..."              # Anthropic API key (or use env ANTHROPIC_API_KEY)
model = "claude-sonnet-4-6"
# api_key_helper = "~/.vibecli/get-key.sh claude"  # Script that prints a fresh key
# thinking_budget_tokens = 10000                    # Enable extended thinking mode

[openai]
enabled = false
api_key = "sk-..."                  # OpenAI API key (or OPENAI_API_KEY)
model = "gpt-4o"
# api_key_helper = "~/.vibecli/get-key.sh openai"

[gemini]
enabled = false
api_key = "AIza..."                 # Google AI Studio key (or GEMINI_API_KEY)
model = "gemini-2.5-flash"
# api_key_helper = "~/.vibecli/get-key.sh gemini"

[grok]
enabled = false
api_key = "..."                     # xAI API key (or GROK_API_KEY)
model = "grok-3-mini"
# api_key_helper = "~/.vibecli/get-key.sh grok"

# ── UI ─────────────────────────────────────────────────────────────

[ui]
theme = "dark"   # "dark" or "light"

# ── Safety ─────────────────────────────────────────────────────────

[safety]
require_approval_for_commands = true      # Prompt before running shell commands
require_approval_for_file_changes = true  # Prompt before applying AI file edits
approval_policy = "suggest"               # "suggest" | "auto-edit" | "full-auto"

# Wildcard tool permission patterns
# denied_tool_patterns = ["bash(rm*)"]   # Block bash calls matching rm*
# denied_tools = ["bash"]               # Exact-match tool block list

# ── Memory (auto-recording) ─────────────────────────────────────────

[memory]
auto_record = false           # Append session learnings to ~/.vibecli/memory.md
min_session_steps = 3         # Minimum tool-use steps before recording triggers

# ── Embedding Index ─────────────────────────────────────────────────

[index]
enabled = true
embedding_provider = "ollama"       # "ollama" or "openai"
embedding_model = "nomic-embed-text"
rebuild_on_startup = false
max_file_size_kb = 500

# ── OpenTelemetry (optional) ─────────────────────────────────────────

[otel]
enabled = false
endpoint = "http://localhost:4318"  # OTLP/HTTP collector
service_name = "vibecli"

# ── Red Team Security Testing ─────────────────────────────────────────

[redteam]
max_depth = 3                      # Max crawl depth for endpoint recon
timeout_secs = 300                 # Per-stage timeout
parallel_agents = 3                # Concurrent exploitation agents
scope_patterns = ["*"]             # URL patterns in scope
exclude_patterns = []              # URLs to skip
auth_config = ""                   # Path to auth YAML file
auto_report = true                 # Auto-generate report on completion

# ── Additional Providers ──────────────────────────────────────────

[groq]
enabled = false
api_key = "gsk_..."                # Groq API key (or GROQ_API_KEY)
model = "llama-3.3-70b-versatile"

[openrouter]
enabled = false
api_key = "sk-or-..."             # OpenRouter key (or OPENROUTER_API_KEY)
model = "anthropic/claude-3.5-sonnet"

[azure_openai]
enabled = false
api_key = "..."                    # Azure key (or AZURE_OPENAI_API_KEY)
api_url = "https://<resource>.openai.azure.com"
model = "gpt-4o"                   # Deployment name

[bedrock]
enabled = false
# Uses AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_REGION env vars
model = "anthropic.claude-3-5-sonnet-20241022-v2:0"

[mistral]
enabled = false
api_key = "..."                    # Mistral key (or MISTRAL_API_KEY)
model = "mistral-large-latest"

[cerebras]
enabled = false
api_key = "..."                    # Cerebras key (or CEREBRAS_API_KEY)
model = "llama3.1-70b"

[deepseek]
enabled = false
api_key = "..."                    # DeepSeek key (or DEEPSEEK_API_KEY)
model = "deepseek-chat"

[zhipu]
enabled = false
api_key = "..."                    # Zhipu key (or ZHIPU_API_KEY)
model = "glm-4"

[vercel_ai]
enabled = false
api_key = "..."                    # Vercel AI key (or VERCEL_AI_API_KEY)
api_url = "https://..."            # Your Vercel AI Gateway URL (REQUIRED)
model = "gpt-4o"

[copilot]
enabled = false
# Uses GITHUB_TOKEN or ~/.config/github-copilot/hosts.json
model = "gpt-4o"

[perplexity]
enabled = false
api_key = "pplx-..."              # Perplexity key (or PERPLEXITY_API_KEY)
model = "sonar-pro"

[together]
enabled = false
api_key = "..."                    # Together AI key (or TOGETHER_API_KEY)
model = "meta-llama/Meta-Llama-3.1-70B-Instruct-Turbo"

[fireworks]
enabled = false
api_key = "..."                    # Fireworks AI key (or FIREWORKS_API_KEY)
model = "accounts/fireworks/models/llama-v3p1-70b-instruct"

[sambanova]
enabled = false
api_key = "..."                    # SambaNova key (or SAMBANOVA_API_KEY)
model = "Meta-Llama-3.1-70B-Instruct"

[minimax]
enabled = false
api_key = "..."                    # MiniMax key (or MINIMAX_API_KEY)
model = "abab6.5s-chat"

# ── Failover (multi-provider chain) ──────────────────────────────

[failover]
chain = ["claude", "openai", "gemini"]   # Try providers in order

# ── Container Sandbox ─────────────────────────────────────────────

[sandbox]
runtime = "docker"                 # "docker", "podman", or "opensandbox"
image = "ubuntu:22.04"
network = false                    # Disable network inside sandbox
memory_limit = "512m"
cpu_limit = "1.0"

# ── Gateway Messaging ─────────────────────────────────────────────

# [[gateway]]
# platform = "telegram"
# bot_token = "..."
# whitelist = ["@username"]

# ── Email (Gmail or Outlook) ──────────────────────────────────────

[email]
provider = "gmail"              # "gmail" or "outlook"
access_token = ""               # OAuth2 access token
# refresh_token = ""            # Optional: used to auto-refresh expired tokens
# default_limit = 20            # Max messages returned per /email inbox call

# ── Calendar (Google Calendar or Outlook) ─────────────────────────

[calendar]
provider = "google"             # "google" or "outlook"
access_token = ""               # OAuth2 access token
# calendar_id = "primary"       # Google Calendar ID (default: "primary")
# timezone = "America/New_York" # Display timezone (default: system timezone)
# default_reminder_minutes = 15 # Add this reminder to every created event
# calendar_readonly = false     # Set true to prevent creates/deletes

# ── Home Assistant ─────────────────────────────────────────────────

[home_assistant]
url = "http://homeassistant.local:8123"   # Base URL of your HA instance
token = ""                                 # Long-lived access token
# insecure = false              # Set true to skip TLS cert verification (self-signed)

# ── Jira ──────────────────────────────────────────────────────────

[jira]
url = "https://yourorg.atlassian.net"   # Jira Cloud or Server base URL
email = ""                               # Atlassian account email
token = ""                               # API token from id.atlassian.com

# ── Notion ────────────────────────────────────────────────────────

notion_api_key = ""             # Notion integration secret (secret_xxx)

# ── Todoist ───────────────────────────────────────────────────────

todoist_api_key = ""            # Todoist API token (Todoist → Settings → Integrations)

Environment Variables

API keys can be set as environment variables instead of (or in addition to) the config file. Environment variables take precedence.

Example:

export ANTHROPIC_API_KEY="sk-ant-..."
vibecli --tui --provider claude

Provider Setup

1. Ollama — Local/Private Models (Default)

Ollama runs models locally on your machine. No API key needed, no data leaves your network.

1. Install Ollama: ollama.ai

2. Pull a coding model:

   ollama pull qwen3-coder              # Default, strong coding
   ollama pull qwen2.5-coder:7b         # Compact, fast
   ollama pull deepseek-coder-v2:16b    # Strong code completion
   ollama pull codellama:13b            # Classic coding model
   

3. Confirm it’s running:

   curl http://localhost:11434/api/tags
   

4. Configure (~/.vibecli/config.toml):

   [ollama]
   enabled = true
   api_url = "http://localhost:11434"
   model = "qwen3-coder:480b-cloud"
   

   Override the base URL with the OLLAMA_HOST env var if Ollama is running on a remote machine.

2. Anthropic Claude — Claude 4 Sonnet/Opus

1. Get an API key at console.anthropic.com

2. Configure:

   [claude]
   enabled = true
   model = "claude-sonnet-4-6"
   # Enable extended thinking (increases latency but improves reasoning):
   # thinking_budget_tokens = 10000
   

   export ANTHROPIC_API_KEY="sk-ant-..."
   

   Available models: claude-opus-4-6 (1M context, highest quality), claude-sonnet-4-6 (200K, balanced), claude-haiku-4-5 (200K, fast/cheap).

   CLI aliases: claude, anthropic

3. OpenAI — GPT-4o and Variants

1. Get an API key at platform.openai.com

2. Configure:

   [openai]
   enabled = true
   model = "gpt-4o"
   

   export OPENAI_API_KEY="sk-..."
   

   Available models: gpt-4o (default), gpt-4o-mini, gpt-4-turbo, o1, o1-mini.

   CLI aliases: openai, gpt

4. Google Gemini — Gemini 2.5 Pro/Flash

1. Get an API key at aistudio.google.com

2. Configure:

   [gemini]
   enabled = true
   model = "gemini-2.5-flash"
   

   export GEMINI_API_KEY="AIza..."
   

   Available models: gemini-2.5-pro (highest quality), gemini-2.5-flash (default, fast), gemini-2.0-flash-lite (cheapest).

   CLI aliases: gemini, google

5. xAI Grok — Grok 2

1. Get an API key at x.ai

2. Configure:

   [grok]
   enabled = true
   model = "grok-3-mini"
   

   export GROK_API_KEY="..."
   

   Available models: grok-2, grok-3-mini (default).

   CLI aliases: grok, xai

6. Groq — Fast Inference (Llama, Mixtral)

Groq runs open-source models on custom LPU hardware with extremely low latency.

1. Get an API key at console.groq.com

2. Configure:

   [groq]
   enabled = true
   model = "llama-3.3-70b-versatile"
   

   export GROQ_API_KEY="gsk_..."
   

   Available models: llama-3.3-70b-versatile (default), llama-3.1-8b-instant, mixtral-8x7b-32768, gemma2-9b-it.

   Free tier available with rate limits.

7. OpenRouter — Multi-Provider Gateway

OpenRouter aggregates 300+ models from multiple providers behind a single API key.

1. Get an API key at openrouter.ai

2. Configure:

   [openrouter]
   enabled = true
   model = "anthropic/claude-3.5-sonnet"
   

   export OPENROUTER_API_KEY="sk-or-..."
   

   Models use the organization/model-name format. Browse available models at openrouter.ai/models.

8. Azure OpenAI — Enterprise Azure-Hosted Models

For organizations using Azure-managed OpenAI deployments with enterprise compliance.

1. Create an Azure OpenAI resource in the Azure Portal

2. Deploy a model (e.g., gpt-4o) and note the deployment name

3. Configure:

   [azure_openai]
   enabled = true
   api_url = "https://<resource>.openai.azure.com"
   model = "gpt-4o"       # Must match your deployment name
   

   export AZURE_OPENAI_API_KEY="..."
   export AZURE_OPENAI_ENDPOINT="https://<resource>.openai.azure.com"
   

   The api_url field is required — it must point to your Azure resource endpoint.

   CLI aliases: azure, azure_openai

9. AWS Bedrock — AWS-Hosted Models (Claude, Llama, Titan)

Uses your existing AWS credentials. No separate API key needed.

1. Enable Bedrock model access in the AWS Console

2. Configure:

   [bedrock]
   enabled = true
   region = "us-east-1"
   model = "anthropic.claude-3-5-sonnet-20241022-v2:0"
   # role_arn = "arn:aws:iam::123456789:role/bedrock-role"  # Optional cross-account
   

   export AWS_ACCESS_KEY_ID="AKIA..."
   export AWS_SECRET_ACCESS_KEY="..."
   export AWS_REGION="us-east-1"
   # export AWS_SESSION_TOKEN="..."  # If using temporary credentials
   

   Available models: anthropic.claude-3-5-sonnet-*, anthropic.claude-3-haiku-*, meta.llama3-*, amazon.titan-*.

   CLI aliases: bedrock, aws, aws-bedrock

10. GitHub Copilot — Copilot Integration

Uses your existing GitHub Copilot subscription.

1. Ensure you have an active GitHub Copilot subscription

2. Configure:

   [copilot]
   enabled = true
   model = "gpt-4o"
   

   Authentication is resolved automatically in this order:

   1. GITHUB_TOKEN env var
   2. ~/.config/github-copilot/hosts.json (from VS Code Copilot extension)
   3. token field in config

   VibeCody automatically exchanges your GitHub token for a short-lived Copilot API token.

   CLI aliases: copilot, github-copilot

11. LocalEdit — Local Code Editing Model

A wrapper around Ollama for local GGUF models. (Note: VibeUI no longer surfaces inline / FIM-style completion — DiffComplete (⌘.) is the only AI editing surface — but this provider remains available for chat and other code paths that route through the standard provider trait.)

   [ollama]
   enabled = true
   api_url = "http://localhost:11434"
   model = "deepseek-coder:6.7b"   # Or any FIM-capable model

No API key required. The model must support fill-in-middle prompting.

12. Mistral — Mistral AI Models

1. Get an API key at console.mistral.ai

2. Configure:

   [mistral]
   enabled = true
   model = "mistral-large-latest"
   

   export MISTRAL_API_KEY="..."
   

   Available models: mistral-large-latest (default), mistral-medium-latest, mistral-small-latest, codestral-latest.

13. Cerebras — Wafer-Scale Inference

Cerebras runs models on CS-3 wafer-scale hardware for ultra-fast inference.

1. Get an API key at cloud.cerebras.ai

2. Configure:

   [cerebras]
   enabled = true
   model = "llama3.1-70b"
   

   export CEREBRAS_API_KEY="..."
   

   Available models: llama3.1-70b (default), llama3.1-8b.

   Free tier available with rate limits.

14. DeepSeek — DeepSeek V3/R1

Strong coding performance at very low prices.

1. Get an API key at platform.deepseek.com

2. Configure:

   [deepseek]
   enabled = true
   model = "deepseek-chat"
   

   export DEEPSEEK_API_KEY="..."
   

   Available models: deepseek-chat (V3, default), deepseek-reasoner (R1, chain-of-thought reasoning).

15. Zhipu — GLM-4 Models

Chinese AI models from Zhipu AI (BigModel).

1. Get an API key at open.bigmodel.cn

2. Configure:

   [zhipu]
   enabled = true
   model = "glm-4"
   

   export ZHIPU_API_KEY="id.secret"   # Format: API ID dot Secret
   

   The API key uses a id.secret format for JWT-based authentication.

   CLI aliases: zhipu, glm

16. Vercel AI — Vercel AI SDK Gateway

A unified proxy that routes to multiple AI services through your Vercel deployment.

1. Deploy a Vercel AI Gateway and get your gateway URL and API key

2. Configure:

   [vercel_ai]
   enabled = true
   api_url = "https://your-gateway.vercel.app/api"   # REQUIRED
   model = "gpt-4o"
   

   export VERCEL_AI_API_KEY="..."
   export VERCEL_AI_GATEWAY_URL="https://your-gateway.vercel.app/api"
   

   The api_url field is required — it must point to your Vercel AI Gateway instance.

   CLI aliases: vercel_ai, vercel

17. MiniMax — MiniMax-Text-01

Chinese AI large language models.

1. Get an API key from minimax.chat

2. Configure:

   [minimax]
   enabled = true
   model = "abab6.5s-chat"
   

   export MINIMAX_API_KEY="..."
   

   Available models: abab6.5s-chat (default), abab6.5-chat, MiniMax-Text-01.

18. Perplexity — Search-Augmented Sonar Models

Combines LLM reasoning with real-time web search. Excellent for research tasks.

1. Get an API key at perplexity.ai

2. Configure:

   [perplexity]
   enabled = true
   model = "sonar-pro"
   

   export PERPLEXITY_API_KEY="pplx-..."
   

   Available models: sonar-pro (default), sonar, sonar-deep-research, sonar-reasoning-pro.

   CLI aliases: perplexity, pplx

19. Together AI — Open Model Hosting (Llama, Qwen)

Hosts open-source models with competitive pricing and a free tier.

1. Get an API key at together.ai

2. Configure:

   [together]
   enabled = true
   model = "meta-llama/Meta-Llama-3.1-70B-Instruct-Turbo"
   

   export TOGETHER_API_KEY="..."
   

   Models use the organization/model-name format. Browse available models at api.together.ai/models.

   CLI aliases: together, together_ai

20. Fireworks AI — Fast Open Model Inference

Fast inference platform for open-source models.

1. Get an API key at fireworks.ai

2. Configure:

   [fireworks]
   enabled = true
   model = "accounts/fireworks/models/llama-v3p1-70b-instruct"
   

   export FIREWORKS_API_KEY="..."
   

   Free tier available with rate limits.

   CLI aliases: fireworks, fireworks_ai

21. SambaNova — Hardware-Accelerated Inference

Runs models on SambaNova’s custom RDU hardware for fast inference.

1. Get an API key at sambanova.ai

2. Configure:

   [sambanova]
   enabled = true
   model = "Meta-Llama-3.1-70B-Instruct"
   

   export SAMBANOVA_API_KEY="..."
   

   Free tier available with rate limits.

apiKeyHelper — Rotating Credentials

For secrets management (Vault, 1Password CLI, AWS Secrets Manager, etc.), use a helper script instead of a static key:

[claude]
# Script is run before each API call; its stdout is used as the Bearer token.
# If the script exits non-zero, the static api_key is used as fallback.
api_key_helper = "~/.vibecli/get-key.sh claude"

Example helper script (~/.vibecli/get-key.sh):

#!/bin/bash
# $1 = provider name
case "$1" in
  claude)  op read "op://Personal/Anthropic/api_key" ;;
  openai)  aws secretsmanager get-secret-value --secret-id openai-key --query SecretString --output text ;;
  *)       echo "" ;;
esac

Selecting a Provider

VibeCLI — at launch:

vibecli --tui --provider ollama          # Default (local)
vibecli --tui --provider claude          # Anthropic Claude
vibecli --tui --provider openai          # OpenAI GPT-4o
vibecli --tui --provider gemini          # Google Gemini
vibecli --tui --provider grok            # xAI Grok
vibecli --tui --provider groq            # Groq (fast Llama/Mixtral)
vibecli --tui --provider openrouter      # OpenRouter gateway
vibecli --tui --provider azure           # Azure OpenAI
vibecli --tui --provider bedrock         # AWS Bedrock
vibecli --tui --provider copilot         # GitHub Copilot
vibecli --tui --provider mistral         # Mistral AI
vibecli --tui --provider cerebras        # Cerebras
vibecli --tui --provider deepseek        # DeepSeek
vibecli --tui --provider zhipu           # Zhipu GLM
vibecli --tui --provider vercel          # Vercel AI Gateway
vibecli --tui --provider minimax         # MiniMax
vibecli --tui --provider perplexity      # Perplexity Sonar
vibecli --tui --provider together        # Together AI
vibecli --tui --provider fireworks       # Fireworks AI
vibecli --tui --provider sambanova       # SambaNova
vibecli --tui --provider failover        # Failover chain

Override the model for any provider:

vibecli --tui --provider claude --model claude-opus-4-6
vibecli --tui --provider openai --model gpt-4o-mini
vibecli --tui --provider groq --model mixtral-8x7b-32768

VibeUI:

Use the provider dropdown in the top bar. The selection is persisted across sessions.

Provider Usage Examples

These examples show real-world workflows across different providers. Each example can be run from the command line or the REPL.

One-Shot Chat

Ask a quick question without entering the REPL:

# Local (free, private)
vibecli --provider ollama "Explain the borrow checker in Rust"

# Cloud (higher quality)
vibecli --provider claude "Explain the borrow checker in Rust"

Agent Mode — Fix a Bug

The agent reads files, edits code, and runs tests autonomously:

# Interactive mode — approve each step
vibecli --agent "Fix the login bug in src/auth.rs" --provider claude

# Example session output:
#   Agent   Fix the login bug in src/auth.rs
#     Policy: suggest (ask before every action)  |  Press Ctrl+C to stop
#
#    ✓ Reading src/auth.rs
#    ✓ Searching: "validate_token"
#
#      bash  Running: cargo check
#       Approve? (y/n/a=approve-all): y
#
#    ✓ Running: cargo check
#    ✓ Patching src/auth.rs (2 hunks)
#    ✓ Running: cargo test auth::tests
#
#   Agent complete: Fixed RS256/HS256 mismatch and hardcoded secret.
#      Files modified: src/auth.rs
#      Commands run: 2
#      Steps: 5/5 succeeded
#      Trace saved: ~/.vibecli/traces/1711234567.jsonl

# Non-interactive — auto-approve everything (CI/scripts)
vibecli --exec "Add error handling to all unwrap() calls in src/" --provider claude --full-auto

# Auto-edit — auto-approve file changes, prompt for shell commands
vibecli --agent "Refactor database module" --provider openai --auto-edit

Agent Mode — Scaffold a Feature

vibecli --agent "Add a REST endpoint POST /api/users that validates email, \
  hashes the password with bcrypt, stores in SQLite, and returns 201" \
  --provider claude --model claude-opus-4-6

Code Review

# Review a file
vibecli --provider claude "Review src/handler.rs for security issues and performance"

# Review a git diff
git diff HEAD~3 | vibecli --provider claude "Review this diff for bugs and suggest improvements"

# Review a GitHub PR
vibecli --provider claude "/review-pr 42"

Multi-Provider Comparison

Send the same prompt to different providers and compare:

# Side-by-side comparison
vibecli --provider claude   "Write FizzBuzz in Rust" > /tmp/claude.txt
vibecli --provider openai   "Write FizzBuzz in Rust" > /tmp/openai.txt
vibecli --provider deepseek "Write FizzBuzz in Rust" > /tmp/deepseek.txt
diff /tmp/claude.txt /tmp/openai.txt

# Or use the Model Arena in the REPL
vibecli
> /arena "Write a binary search in Python" --providers claude,openai,gemini

REPL — Interactive Session

vibecli
> What does this function do?
> [src/auth.rs]             # Attach a file for context
> Now add rate limiting to the validate_token function
> /model claude-opus-4-6    # Switch to a more capable model mid-session
> /cost                     # Check session token costs
> /quit

Provider-Specific Strengths

Ollama — Fully offline, unlimited use:

# Pull a model and chat immediately (no API key, no cost)
ollama pull qwen3-coder
vibecli --provider ollama --model qwen3-coder "Optimize this SQL query: SELECT * FROM users WHERE ..."

# Remote Ollama server
OLLAMA_HOST=http://gpu-server:11434 vibecli --provider ollama "Explain this stack trace"

Claude — Extended thinking for hard problems:

# ~/.vibecli/config.toml
[claude]
enabled = true
model = "claude-sonnet-4-6"
thinking_budget_tokens = 16000   # AI "thinks" internally before answering

# Architecture decisions, complex debugging, security analysis
vibecli --provider claude --model claude-opus-4-6 \
  "Design a migration strategy from our monolith to microservices. \
   Here is the current architecture:" --add-dir ./src/

Gemini — Massive context window:

# Feed an entire codebase (Gemini supports up to 2M tokens)
vibecli --provider gemini --model gemini-2.5-pro \
  "Summarize the architecture of this project and identify dead code" \
  --add-dir ./src/

Groq / Cerebras / SambaNova — Ultra-fast inference:

# Near-instant responses for quick iterations
vibecli --provider groq "Convert this JSON to a Rust struct: {\"name\": \"Alice\", \"age\": 30}"
vibecli --provider cerebras "Write a regex to match email addresses"
vibecli --provider sambanova "Explain this error: cannot borrow as mutable"

DeepSeek — Budget-friendly coding:

# Strong coding quality at ~1/10th the price of Claude/GPT-4o
vibecli --provider deepseek "Write comprehensive unit tests for src/auth.rs"
vibecli --provider deepseek --model deepseek-reasoner "Debug this race condition" # R1 reasoning

Perplexity — Search-augmented answers:

# Answers grounded in real-time web search
vibecli --provider perplexity "What are the latest breaking changes in Tokio 1.40?"
vibecli --provider perplexity "Compare axum vs actix-web performance benchmarks 2026"

OpenRouter — Access any model:

# Try niche or new models via unified API
vibecli --provider openrouter --model "meta-llama/llama-3.3-70b" "Explain monads"
vibecli --provider openrouter --model "google/gemini-2.5-pro" "Review this code"
vibecli --provider openrouter --model "deepseek/deepseek-r1" "Solve this optimization"

Failover — Automatic resilience:

# ~/.vibecli/config.toml — tries each provider in order
[failover]
chain = ["claude", "openai", "gemini", "ollama"]

# If Claude is rate-limited, automatically falls back to OpenAI, then Gemini, then local Ollama
vibecli --provider failover --agent "Fix the build errors in src/"

Enterprise Examples

Azure OpenAI — Corporate proxy with compliance:

export AZURE_OPENAI_API_KEY="..."
export AZURE_OPENAI_ENDPOINT="https://mycompany.openai.azure.com"
vibecli --provider azure --model gpt-4o "Audit this code for OWASP top 10 vulnerabilities"

AWS Bedrock — IAM-authenticated, no API keys in code:

# Uses your AWS CLI credentials or instance role
aws sso login --profile production
vibecli --provider bedrock "Generate a CloudFormation template for an ECS Fargate service"

GitHub Copilot — Use your existing subscription:

# No extra API key needed if you have VS Code Copilot configured
vibecli --provider copilot "Complete this function" --file src/parser.rs

Batch Processing & CI/CD

# One-shot in CI pipelines
vibecli --exec "Check this diff for security issues" --provider claude --full-auto < pr.diff

# Process multiple files
for f in src/*.rs; do
  vibecli --provider groq "Summarize what this file does" --file "$f" >> summary.md
done

# JSON output for scripts
vibecli --provider claude --json "List all TODO comments in src/" | jq '.items[]'

Vision (Image Analysis)

# Analyze screenshots, diagrams, or UI mockups (Claude, OpenAI, Gemini)
vibecli --provider claude "What's wrong with this UI?" --image ./screenshot.png
vibecli --provider gemini "Convert this wireframe to React components" --image ./mockup.png

Cost Tracking

vibecli
> /cost
Session cost summary:
  claude:   $0.0342 (12,400 tokens)
  openai:   $0.0128 (5,200 tokens)
  ollama:   $0.0000 (8,100 tokens)  [local]
  Total:    $0.0470

Safety Settings

VibeCLI has a built-in approval gate for potentially destructive actions.

Approval Policies

Override at launch: vibecli --full-auto, vibecli --auto-edit, vibecli --suggest

Wildcard Tool Patterns

Block granular tool+argument combinations without blocking the entire tool:

[safety]
# Block bash calls matching rm* but allow everything else
denied_tool_patterns = ["bash(rm*)", "bash(sudo*)"]

# Still block specific tools entirely
denied_tools = ["execute_python"]

Disable approvals for trusted environments:

[safety]
require_approval_for_commands = false
require_approval_for_file_changes = false
approval_policy = "full-auto"

Warning: Disabling command approval allows VibeCLI to execute AI-suggested commands without confirmation. Only disable in trusted, sandboxed environments.

Auto Memory Recording

VibeCLI can automatically summarize completed agent sessions and append key learnings to ~/.vibecli/memory.md.

[memory]
auto_record = true         # Enable auto-recording (default: false)
min_session_steps = 3      # Only record sessions with ≥ N tool calls

After a session completes, the LLM generates 1–3 concise bullet points and appends them:

<!-- auto-recorded 2026-02-24 -->
- Always check for existing tests before adding new ones
- Use `cargo check` before `cargo build` to catch compile errors faster

The memory file is automatically injected into future agent system prompts.

Tip: Keep memory.md under ~2 000 tokens. For long-lived or large knowledge sets, use the OpenMemory cognitive store (see below) — it retrieves only what is relevant rather than injecting everything.

OpenMemory Cognitive Engine

OpenMemory is a persistent, structured memory store with five cognitive sectors, a temporal knowledge graph, and an optional verbatim drawer layer (MemPalace techniques). It is the recommended memory system for anything beyond short-term session notes.

Configuration

[openmemory]
enabled = true              # Enable the cognitive memory engine (default: true)
auto_inject = true          # Inject context into every agent turn (default: true)
max_context_tokens = 1200   # Hard cap on injected context tokens (default: 1200)
decay_enabled = true        # Run salience decay each session (default: true)
consolidate_on_exit = false # Run sleep-cycle consolidation when the REPL exits
encryption = false          # AES-256-GCM at rest — run /openmemory encrypt to enable

# Verbatim drawer layer (MemPalace techniques)
drawer_chunk_size = 800     # Characters per verbatim chunk (default: 800)
drawer_overlap = 100        # Overlap between adjacent chunks (default: 100)
drawer_dedup_threshold = 0.85   # Cosine near-dedup threshold (default: 0.85)

# 4-layer context tuning
l1_tokens = 700             # Token budget for Essential Story (L1) layer
l2_limit = 8                # Max results from scoped semantic search (L2)
l3_threshold = 3            # Trigger L3 (verbatim drawers) when L2 < this many results

Storage paths

Each store contains four files: memories.json, waypoints.json, facts.json, drawers.json.

Encryption

To enable AES-256-GCM encryption at rest, run /openmemory encrypt in the REPL or set encryption = true in config.toml. The key is stored at <store>/.key (mode 0600).

To use a passphrase instead of a stored key:

VIBECLI_MEMORY_KEY="$(pass show vibecli/memory)" vibecli

Tuning decay rates

Each sector has an independent decay rate (fraction of salience lost per day when a memory is not accessed):

[openmemory.decay]
episodic   = 0.015   # ~67 days to half-salience
semantic   = 0.005   # ~200 days to half-salience
procedural = 0.008   # ~125 days to half-salience
emotional  = 0.020   # ~50 days to half-salience
reflective = 0.001   # ~1000 days to half-salience (very persistent)
purge_threshold = 0.05  # Remove memories below 5% salience

Disabling memory injection for a session

vibecli --no-memory                    # Skip injection entirely
vibecli --memory-scope ./my-project    # Use project-scoped store only
vibecli --dry-run-memory "query"       # Print what would be injected, then exit

Rules Directory

Place .md files in .vibecli/rules/ (project-level) or ~/.vibecli/rules/ (global) to inject persistent instructions into the agent system prompt.

Rules support optional YAML-style front-matter to restrict injection to specific file types:

name: rust-safety
path_pattern: "**/*.rs"
When editing Rust files, prefer `?` over `unwrap()`.
Always add `#[must_use]` to functions returning Result or Option.

Rules without a path_pattern always inject. Rules with a pattern only inject when a matching file is open in the session’s workspace.

Productivity Integrations

VibeCLI connects to external services for email, calendar, tasks, knowledge, project tracking, and smart home control. Configure each under the corresponding section in config.toml or via environment variables.

Email (Gmail / Outlook)

[email]
provider     = "gmail"    # "gmail" | "outlook"
access_token = "ya29.xxxx"

Getting a Gmail token: Use the Google OAuth2 Playground (developers.google.com/oauthplayground) with scope https://www.googleapis.com/auth/gmail.modify, or run vibecli --setup --email.

Getting an Outlook token: Use the Microsoft OAuth2 flow with scope https://graph.microsoft.com/Mail.ReadWrite.

Calendar (Google / Outlook)

[calendar]
provider     = "google"   # "google" | "outlook"
access_token = "ya29.xxxx"
timezone     = "America/New_York"

Home Assistant

[home_assistant]
url   = "http://homeassistant.local:8123"
token = "eyJ0..."  # Settings → Profile → Long-Lived Access Tokens

Todoist

todoist_api_key = "xxxx"   # Todoist → Settings → Integrations → API token

Notion

notion_api_key = "secret_xxxx"   # notion.so/my-integrations

Share pages/databases with the integration in Notion before they appear in search results.

Jira

[jira]
url   = "https://yourorg.atlassian.net"
email = "you@yourorg.com"
token = "ATATT3xxx"   # id.atlassian.com → Security → API tokens

Command History

VibeCLI saves REPL command history to ~/.vibecli/history.txt. This file is created automatically and persists across sessions. History is navigable with the Up/Down arrow keys in the REPL.