Configuration

ByteMind loads config from the first path it finds, in this order:

1. Path passed via -config <path> flag
2. .bytemind/config.json in the current working directory
3. config.json in the current working directory
4. ~/.bytemind/config.json in your home directory

OpenAI-Compatible Providers

Works with OpenAI, DeepSeek, Azure OpenAI, and any service that implements the /v1/chat/completions interface:

{
  "provider": {
    "type": "openai-compatible",
    "base_url": "https://api.openai.com/v1",
    "model": "gpt-4o",
    "api_key": "YOUR_API_KEY"
  },
  "approval_policy": "on-request",
  "max_iterations": 32,
  "stream": true
}

Pass the API key via environment variable (recommended 閳?keeps secrets out of files):

{
  "provider": {
    "type": "openai-compatible",
    "base_url": "https://api.openai.com/v1",
    "model": "gpt-4o",
    "api_key_env": "OPENAI_API_KEY"
  }
}

export OPENAI_API_KEY="sk-..."
bytemind chat

Anthropic

{
  "provider": {
    "type": "anthropic",
    "base_url": "https://api.anthropic.com",
    "model": "claude-sonnet-4-20250514",
    "api_key": "YOUR_API_KEY",
    "anthropic_version": "2023-06-01"
  },
  "approval_policy": "on-request",
  "max_iterations": 32,
  "stream": true
}

Local / Custom Models

Any endpoint that speaks the OpenAI chat completions format works:

{
  "provider": {
    "type": "openai-compatible",
    "base_url": "http://localhost:11434/v1",
    "model": "qwen2.5-coder:7b",
    "api_key": "ollama"
  }
}

Auto-detect provider type

Set "auto_detect_type": true to let ByteMind infer the provider type from base_url automatically.

Approval Policy

approval_policy controls when high-risk tools (file writes, shell commands) pause for confirmation:

Value	Behavior
on-request (default)	Agent waits for confirmation before each high-risk operation

approval_mode sets the overall interaction style:

Value	Behavior
interactive (default)	Prompt for approval on each operation
full_access	Auto-approve approval-required actions with no prompt

Legacy compatibility is now gated: approval_mode: away is blocked by default to prevent silent privilege escalation. Set BYTEMIND_ALLOW_AWAY_FULL_ACCESS=true only as a temporary migration escape hatch.

away_policy (deprecated compatibility field):

Value	Behavior
auto_deny_continue (default)	Accepted for compatibility; no runtime behavior change
fail_fast	Accepted for compatibility; no runtime behavior change

Full access caution

In full_access mode, approval-required operations are auto-approved with no prompt. Keep sandbox/allowlist settings strict when running unattended.

Sandbox

When sandbox is enabled, file and shell tools are restricted to the declared writable roots:

{
  "sandbox_enabled": true,
  "writable_roots": ["./src", "./tests"]
}

You can also enable it via environment variables:

BYTEMIND_SANDBOX_ENABLED=true BYTEMIND_WRITABLE_ROOTS=./src bytemind chat

Iteration Budget

max_iterations caps the number of tool-call rounds per task, preventing runaway loops:

{
  "max_iterations": 64
}

When the limit is reached, the agent produces a progress summary and stops gracefully 閳?it does not error out. Raise this value for complex refactors or large migrations.

Token Quota

token_quota sets the warning threshold for token consumption per task (default: 300,000):

{
  "token_quota": 500000
}

Full Example

{
  "provider": {
    "type": "openai-compatible",
    "base_url": "https://api.openai.com/v1",
    "model": "gpt-4o",
    "api_key_env": "OPENAI_API_KEY"
  },
  "approval_policy": "on-request",
  "approval_mode": "interactive",
  "max_iterations": 32,
  "stream": true,
  "update_check": { "enabled": true },
  "context_budget": {
    "warning_ratio": 0.85,
    "critical_ratio": 0.95,
    "max_reactive_retry": 1
  }
}

See Config Reference for the full field list.