Configuration Reference

OpenParallax is configured through a config.yaml file in your workspace directory (typically ~/.openparallax/<agent-name>/config.yaml). This file is created by the openparallax init wizard and can be edited manually at any time.

Sources: internal/types/config.go, internal/config/defaults.go

Top-Level

Key	Type	Default	Description
workspace	string	"."	Root directory for the agent's workspace files

Models (models)

The model pool defines all available LLM providers. This is a list of model entries.

Key	Type	Default	Description
models[].name	string	—	Unique identifier for this model (e.g. "claude-sonnet")
models[].provider	string	—	LLM provider: "anthropic", "openai", "google", "ollama"
models[].model	string	—	Provider-specific model identifier
models[].api_key_env	string	—	Environment variable holding the API key
models[].base_url	string	—	Override the provider's default API endpoint

Roles (roles)

Maps functional roles to model names from the model pool.

Key	Type	Default	Description
roles.chat	string	—	Model name for the main conversation
roles.shield	string	—	Model name for Tier 2 Shield evaluation
roles.embedding	string	—	Model name for vector embeddings
roles.sub_agent	string	—	Model name for sub-agent tasks
roles.image	string	—	Model name for image generation
roles.video	string	—	Model name for video generation

Shield (shield)

Configures the 4-tier Shield security pipeline. See Security for architecture details.

Key	Type	Default	Description
shield.policy_file	string	"security/shield/default.yaml"	Path to the YAML policy file (Tier 0). Read-only via API
shield.onnx_threshold	float64	0.85	Confidence threshold for the ONNX classifier (0.0-1.0)
shield.heuristic_enabled	bool	true	Enable the heuristic regex classifier (Tier 1)
shield.classifier_enabled	bool	false	Enable the ML classifier. Requires classifier_mode: sidecar and a running sidecar binary. Default is heuristic-only (79 rules).
shield.classifier_mode	string	"sidecar"	How the ONNX classifier runs when enabled. "sidecar" connects to an external classifier service via classifier_addr. This is the only supported mode.
shield.classifier_addr	string	—	Address of the ONNX classifier sidecar when classifier_mode: sidecar (e.g. "localhost:8090")
shield.classifier_skip_types	[]string	see below	Action types where ONNX classification is bypassed because the trained model over-fires on benign payloads. Heuristics and policy rules still run. Default: [write_file, delete_file, move_file, copy_file, send_email, send_message, http_request]. See Shield Tier 1 → Per-Action-Type ONNX Skip List.

Shield Evaluator

The Tier 2 evaluator is configured via roles.shield, which maps to a model entry in the models[] pool. There is no shield.evaluator config block. Use AgentConfig.ShieldModel() to resolve the evaluator model programmatically.

Shield Tier 3 (shield.tier3)

Human-in-the-loop approval for uncertain Shield verdicts.

Key	Type	Default	Description
shield.tier3.max_per_hour	int	10	Maximum Tier 3 prompts per hour
shield.tier3.timeout_seconds	int	300	Seconds to wait for user response before auto-deny

Identity (identity)

Key	Type	Default	Description
identity.name	string	"Atlas"	Agent display name
identity.avatar	string	—	Emoji displayed alongside the agent name

Channels (channels)

Configures messaging platform adapters. Each channel is optional and enabled independently.

WhatsApp (channels.whatsapp)

Key	Type	Default	Description
channels.whatsapp.enabled	bool	false	Enable the WhatsApp adapter
channels.whatsapp.phone_number_id	string	—	WhatsApp Business phone number ID
channels.whatsapp.access_token_env	string	—	Environment variable for the access token
channels.whatsapp.verify_token	string	—	Webhook verification token
channels.whatsapp.webhook_port	int	—	Webhook listen port
channels.whatsapp.allowed_numbers	[]string	—	Allowlist of phone numbers

Telegram (channels.telegram)

Key	Type	Default	Description
channels.telegram.enabled	bool	false	Enable the Telegram adapter
channels.telegram.token_env	string	—	Environment variable for the bot token
channels.telegram.allowed_users	[]int64	—	Allowlist of Telegram user IDs
channels.telegram.allowed_groups	[]int64	—	Allowlist of Telegram group IDs
channels.telegram.private_only	*bool	—	When true, ignore all group messages
channels.telegram.polling_interval	int	—	Polling interval in seconds

Discord (channels.discord)

Key	Type	Default	Description
channels.discord.enabled	bool	false	Enable the Discord adapter
channels.discord.token_env	string	—	Environment variable for the bot token
channels.discord.allowed_guilds	[]string	—	Allowlist of Discord server (guild) IDs
channels.discord.allowed_channels	[]string	—	Allowlist of Discord channel IDs
channels.discord.allowed_users	[]string	—	Allowlist of Discord user IDs
channels.discord.respond_to_mentions	bool	false	Respond when the bot is @mentioned

Slack (channels.slack)

Key	Type	Default	Description
channels.slack.enabled	bool	false	Enable the Slack adapter
channels.slack.bot_token_env	string	—	Environment variable for the bot OAuth token
channels.slack.app_token_env	string	—	Environment variable for the app-level token

Signal (channels.signal)

Key	Type	Default	Description
channels.signal.enabled	bool	false	Enable the Signal adapter
channels.signal.cli_path	string	—	Path to the signal-cli binary
channels.signal.account	string	—	Signal account phone number
channels.signal.allowed_numbers	[]string	—	Allowlist of phone numbers

Teams (channels.teams)

Key	Type	Default	Description
channels.teams.enabled	bool	false	Enable the Microsoft Teams adapter
channels.teams.app_id_env	string	—	Environment variable for the Teams app ID
channels.teams.password_env	string	—	Environment variable for the Teams app password

iMessage (channels.imessage)

macOS only.

Key	Type	Default	Description
channels.imessage.enabled	bool	false	Enable the iMessage adapter
channels.imessage.apple_id	string	—	Apple ID email used in Messages.app

Chronicle (chronicle)

Configures copy-on-write workspace snapshots for state versioning and rollback.

Key	Type	Default	Description
chronicle.max_snapshots	int	100	Maximum number of snapshots to retain
chronicle.max_age_days	int	30	Maximum age of snapshots in days

Web (web)

Configures the Web UI server.

Key	Type	Default	Description
web.enabled	bool	true	Enable the Web UI
web.host	string	""	Bind address. Empty resolves to 127.0.0.1 (localhost only). Set to "0.0.0.0" for remote access (requires password_hash)
web.port	int	3100	HTTP listen port. Read-only via API
web.grpc_port	int	0 (dynamic)	gRPC listen port for CLI-Engine communication
web.auth	bool	true	Enable cookie-based authentication
web.password_hash	string	—	Bcrypt hash of the web UI password. Required when host is non-localhost
web.allowed_origins	[]string	—	Origins permitted for CORS and WebSocket. Empty = localhost only

Agents (agents)

Configures sub-agent orchestration, LLM call limits, tool timeouts, and crash recovery.

Key	Type	Default	Description
agents.sub_agent_model	string	—	Override the default sub-agent model. Empty = auto-detect cheapest
agents.max_rounds	int	—	Max LLM calls per sub-agent (default 20)
agents.max_tool_rounds	int	25	Max tool-call round-trips per message
agents.context_window	int	128000	Assumed model context window in tokens
agents.compaction_threshold	int	70	Context budget percentage (0-100) that triggers both compaction and the history/current-turn split inside the compactor
agents.max_response_tokens	int	4096	Max tokens per LLM response
agents.shell_timeout_seconds	int	30	Default shell command timeout. Increase for long builds or large git operations
agents.browser_nav_timeout_seconds	int	30	Browser page navigation timeout. Increase for slow pages on poor connections
agents.browser_idle_minutes	int	5	Minutes of inactivity before the headless browser session is shut down to free memory
agents.sub_agent_timeout_seconds	int	900	Max time a sub-agent can run before being killed (default 15 minutes). Override per-spawn via the timeout parameter on create_agent
agents.max_concurrent_sub_agents	int	10	Cap on simultaneously running sub-agent processes. Once reached, create_agent returns an error until one completes. Raise it if you intentionally orchestrate large fan-outs
agents.max_sub_agent_rounds	int	20	Maximum number of LLM calls each sub-agent can make before stopping
agents.crash_restart_budget	int	5	Max agent (or engine) crashes within crash_window_seconds before the process manager stops restarting
agents.crash_window_seconds	int	60	Time window for crash counting
agents.max_consecutive_nav_failures	int	3	Consecutive browser navigation failures before the executor disables navigation for the session. Prevents wasted LLM round-trips on hosts where the browser fundamentally cannot load pages (e.g. Flatpak sandbox)

General (general)

Global operational settings.

Key	Type	Default	Description
general.fail_closed	bool	true	All Shield evaluation errors result in BLOCK
general.rate_limit	int	30	Maximum actions per minute
general.verdict_ttl_seconds	int	60	How long a Shield verdict remains valid
general.daily_budget	int	100	Maximum Tier 2 LLM evaluator calls per day. Read-only via API
general.output_sanitization	bool	false	Wrap tool results in data boundaries to mitigate prompt injection. Increases token usage slightly

Security (security)

Security subsystem policy paths. The subsystems are non-negotiable; only the policies are tunable.

Key	Type	Default	Description
security.ifc_policy	string	"security/ifc/default.yaml"	Path to the IFC policy YAML file. Three presets ship: default, permissive, strict. See IFC reference
security.override_mode	string	""	Overrides the mode declared in the IFC policy. Empty = use policy's own mode. "audit" = log but don't block. "enforce" = block. Not in SettableKeys — requires restart
security.memory_block_levels	[]string	[]	Sensitivity levels that block memory writes when session taint reaches them. Overridden by memory_block_levels in the IFC policy file if present. Empty = use the IFC policy's setting (or built-in default [critical, restricted]). See IFC reference

Embedding

The embedding provider is configured via roles.embedding, which maps to a model entry in the models[] pool. There is no memory top-level config key. Use AgentConfig.EmbeddingModel() to resolve the embedding model programmatically.

MCP (mcp)

External Model Context Protocol server connections. Each server runs as a child process and provides additional tools.

Key	Type	Default	Description
mcp.servers	[]MCPServerConfig	—	List of MCP servers to connect
mcp.servers[].name	string	—	Server display name
mcp.servers[].command	string	—	Command to start the server
mcp.servers[].args	[]string	—	Command arguments
mcp.servers[].env	map[string]string	—	Environment variables for the server process

Email (email)

Configures email sending (SMTP) and reading (IMAP).

Key	Type	Default	Description
email.provider	string	—	Email provider ("smtp" for now)

SMTP (email.smtp)

Key	Type	Default	Description
email.smtp.host	string	—	SMTP server hostname
email.smtp.port	int	—	SMTP server port
email.smtp.username	string	—	SMTP login username
email.smtp.password	string	—	SMTP login password
email.smtp.from	string	—	Sender email address
email.smtp.tls	bool	—	Enable TLS encryption

IMAP (email.imap)

Key	Type	Default	Description
email.imap.host	string	—	IMAP server hostname (e.g. "imap.gmail.com")
email.imap.port	int	—	IMAP server port (typically 993 for TLS)
email.imap.tls	bool	—	Enable TLS encryption
email.imap.username	string	—	IMAP login username (for password auth)
email.imap.password	string	—	IMAP login password or app password (for password auth)
email.imap.auth_mode	string	—	Authentication mode: "password" or "oauth2"
email.imap.account	string	—	Email address for OAuth2 token lookup

Calendar (calendar)

Key	Type	Default	Description
calendar.provider	string	—	Calendar provider: "google", "caldav", "microsoft"
calendar.google_credentials_file	string	—	Path to Google OAuth credentials file
calendar.calendar_id	string	—	Google calendar ID
calendar.caldav_url	string	—	CalDAV server URL
calendar.caldav_username	string	—	CalDAV username
calendar.caldav_password	string	—	CalDAV password
calendar.microsoft_account	string	—	Microsoft account email for OAuth2 token lookup

OAuth (oauth)

OAuth2 client credentials for email and calendar integrations. Tokens are obtained via openparallax auth <provider>.

Google (oauth.google)

Key	Type	Default	Description
oauth.google.client_id	string	—	Google OAuth2 application client ID
oauth.google.client_secret	string	—	Google OAuth2 application client secret

Microsoft (oauth.microsoft)

Key	Type	Default	Description
oauth.microsoft.client_id	string	—	Microsoft OAuth2 application client ID
oauth.microsoft.client_secret	string	—	Microsoft OAuth2 application client secret
oauth.microsoft.tenant_id	string	"common"	Azure AD tenant ID

Tools (tools)

Key	Type	Default	Description
tools.disabled_groups	[]string	—	Tool group names to disable. See Action Types — Tool Groups

Skills (skills)

Key	Type	Default	Description
skills.disabled	[]string	—	Skill names to disable

Read-Only Fields

The following fields cannot be changed through the web UI settings API. Edit config.yaml directly and restart:

• web.port
• general.daily_budget (shield.tier2_budget in the API)
• shield.policy_file

Minimal Example

workspace: /home/user/.openparallax/atlas

models:
  - name: claude-sonnet
    provider: anthropic
    model: claude-sonnet-4-6
    api_key_env: ANTHROPIC_API_KEY
  - name: claude-haiku
    provider: anthropic
    model: claude-haiku-4-5-20251001
    api_key_env: ANTHROPIC_API_KEY

roles:
  chat: claude-sonnet
  shield: claude-haiku

identity:
  name: Atlas

shield:
  policy_file: security/shield/default.yaml
  heuristic_enabled: true

web:
  enabled: true
  port: 3100

general:
  fail_closed: true
  rate_limit: 30
  daily_budget: 100