Channel aliases this agent handles (e.g. ["telegram.<alias>", "discord.<alias>"]). Each entry is a ChannelRef resolving through [channels.<type>.<alias>]; Config::validate() fails loud on dangling references.

Set it on any surface:

zeroclaw config set agents.<alias>.channels <value>

export ZEROCLAW_agents__<alias>__channels=

Optional override for the per-message LLM reply-intent classifier (classify_channel_reply_intent in zeroclaw-channels). When non-empty, the channel orchestrator routes the “should this message be replied to?” classification call to [providers.models.<type>.<alias>] referenced here, instead of reusing the main agent’s model_provider. Source of truth for api_key / uri / model / temperature etc. is the referenced [providers.models.<type>.<alias>] entry. This field is a reference only (NEVER a copy), per AGENTS.md SINGLE SOURCE OF TRUTH. Empty (Default) = inherit the main agent’s resolved provider+model (preserves pre-PR behavior; backward compatible). Use case: classification is a cheap REPLY/NO_REPLY decision, doesn’t need a high-end model. Point this at a fast/free small model (e.g. kimi-k2.5, qwen-turbo) while model_provider stays on the expensive answering model (e.g. qwen3.6-plus). Note: TOML table names cannot contain ., so alias kimi-k2.5 must be written as [providers.models.custom.kimi-k2-5]. The underlying model = "kimi-k2.5" string can still contain dots. ACP channels (IDE-direct) always reply and skip the classifier entirely, so this field has no effect on ACP traffic.

Set it on any surface:

zeroclaw config set agents.<alias>.classifier_provider <value>

export ZEROCLAW_agents__<alias>__classifier_provider=

Cron job aliases. Each entry references cron[key], a declarative scheduled job invoked by the scheduler on its configured trigger. When the cron fires, this agent is the actor that executes the job.

Set it on any surface:

zeroclaw config set agents.<alias>.cron_jobs <value>

export ZEROCLAW_agents__<alias>__cron_jobs=

Per-agent identity format ([agents.<alias>.identity]). Each agent renders its own IDENTITY.md / SOUL.md inside its per-agent workspace; this block selects the format (OpenClaw or AIEOS) and optional inline/file source for the agent’s identity document.

Set it on any surface:

zeroclaw config set agents.<alias>.identity <value>

export ZEROCLAW_agents__<alias>__identity=

Knowledge bundle aliases. Additive: the agent loads every listed bundle.

Set it on any surface:

zeroclaw config set agents.<alias>.knowledge_bundles <value>

export ZEROCLAW_agents__<alias>__knowledge_bundles=

MCP bundle aliases. Each entry references mcp_bundles[key], itself a named group of MCP servers; agents pick which bundles to load.

Set it on any surface:

zeroclaw config set agents.<alias>.mcp_bundles <value>

export ZEROCLAW_agents__<alias>__mcp_bundles=

Per-agent memory backend selection ([agents.<alias>.memory]). The backend field is locked at agent creation and immutable on subsequent loads. Defaults to Sqlite. See crate::multi_agent::AgentMemoryConfig.

Set it on any surface:

zeroclaw config set agents.<alias>.memory <value>

export ZEROCLAW_agents__<alias>__memory=

Dotted model-provider alias (e.g. "anthropic.<alias>"). Resolves through model_providers.<type>.<alias> at runtime; Config::validate() fails loud on dangling references.

Set it on any surface:

zeroclaw config set agents.<alias>.model_provider <value>

export ZEROCLAW_agents__<alias>__model_provider=

Risk profile alias (e.g. "default"). Resolves delegation guardrails at runtime.

Set it on any surface:

zeroclaw config set agents.<alias>.risk_profile <value>

export ZEROCLAW_agents__<alias>__risk_profile=

Runtime profile alias (e.g. "default"). Resolves agentic/iteration settings.

Set it on any surface:

zeroclaw config set agents.<alias>.runtime_profile <value>

export ZEROCLAW_agents__<alias>__runtime_profile=

Skill bundle aliases. Each entry resolves to skill_bundles[key].directory at runtime; the agent loads every listed bundle.

Set it on any surface:

zeroclaw config set agents.<alias>.skill_bundles <value>

export ZEROCLAW_agents__<alias>__skill_bundles=

Transcription / STT provider as a dotted alias reference (<type>.<alias>, e.g. "groq.<alias>"). Resolves through transcription_providers.<type>.<alias>. Empty = agent has no transcription preference; channels that ingest voice still need a resolved provider (there is no global default), so an inbound voice flow into an agent with empty transcription_provider errors loudly at the channel boundary.

Set it on any surface:

zeroclaw config set agents.<alias>.transcription_provider <value>

export ZEROCLAW_agents__<alias>__transcription_provider=

TTS provider as a dotted alias reference (<type>.<alias>, e.g. "openai.<alias>"). Resolves through tts_providers.<type>.<alias>. Empty = no TTS for this agent (there is no global default-provider concept; every agent that wants TTS sets its own tts_provider).

Set it on any surface:

zeroclaw config set agents.<alias>.tts_provider <value>

export ZEROCLAW_agents__<alias>__tts_provider=

Per-agent workspace block ([agents.<alias>.workspace]). Holds the agent’s filesystem path, cross-agent access allowlist, filesystem-escape boolean, and cross-agent memory allowlist. Default is fully jailed (no cross-agent access). See crate::multi_agent::AgentWorkspaceConfig.

Set it on any surface:

zeroclaw config set agents.<alias>.workspace <value>

export ZEROCLAW_agents__<alias>__workspace=