Agents - cliq

An agent is a registered runtime the daemon invokes for a phase (agent: <name> in the team manifest). Built-ins ship with the platform and are classified by agent_type. For how agents fit into teams and when roles apply, see Concepts. To run a team end-to-end, see Running teams. Configure credentials with cliq setup or:

cliq settings agents.<name>.<key> <value>

Agent types

Type	Role?	Typical inputs	Used for
`llm`	Yes	`role`, `model`, …	Model / coding-CLI work and LLM gates
`exec`	No	`commands`	Deterministic shell steps
`connector`	No	`action`, `sources`, `target_entries`	External systems
`gate`	Special	Gate / review config	`hug`, `auto-gate`
`notify`	No	notify channel wiring	Slack, email, …
`meta`	No	`team`	Nested team launch

LLM agents

Shared phase inputs (all agent_type: llm):

Input	Required	Default	Description
`role`	yes	—	Specialist brief for this phase (prompt text)
`model`	no	per agent	Model id
`temperature`	no	—	Sampling temperature
`max_turns`	no	—	Max conversation turns
`needs_filesystem`	no	varies	Whether the coding tool may use the workspace
`task`	no	runtime	Requirement / instruction (often from run inputs)

Capabilities: phase:standard, phase:gate (unless noted).

Agent	Settings (required)	Default model	Notes
`cursor`	`api_key`	`claude-sonnet-4-20250514`	Coding CLI (`agent` binary); `needs_filesystem` default `true`
`claude-code`	`api_key`	`claude-sonnet-4-20250514`	Coding CLI (`claude`); `needs_filesystem` default `true`
`gemini`	`api_key`	`gemini-2.5-pro`	Coding CLI (`gemini`); `needs_filesystem` default `true`
`codex`	`api_key`	`gpt-4.1`	Coding CLI (`codex`); `needs_filesystem` default `true`
`claude-api`	`api_key`	`claude-sonnet-4-20250514`	Anthropic REST API
`gemini-api`	`api_key`	`gemini-2.5-flash`	Google API
`openai-api`	`api_key`	`gpt-4o`	OpenAI API

Example phase:

- name: implement
  type: standard
  agent: cursor
  depends_on: [plan]

Role text for the phase is supplied when the team is assembled / run (LLM phases only). Model overrides can be set on the phase payload where the team author allows it.

Exec agents

Agent	Inputs	Description
`exec`	`commands` (`command[]`, required) — each `{ run: string }`	Run shell commands in order
`echo`	(none)	Echo upstream data (test / plumbing)

- name: greet
  type: standard
  agent: exec
  commands:
    - run: echo "hello from exec"

Connector agents

Connectors talk to external systems. They do not use roles. Common fields:

Input	Required	Description
`action`	often yes	Operation to perform (agent-specific)
`sources`	no*	Resources to fetch (`source_entry[]`)
`target_entries`	no*	Resources to write (`target_entry[]`)

*For pull/push agents, at least one of sources / target_entries is typically required.

Agent	Settings (required)	Inputs	Capabilities
`curl`	—	`sources`, `target_entries`	`phase:standard`
`gdrive`	`credentials_file`	`sources`, `target_entries`	`phase:standard`
`s3`	`access_key_id`, `secret_access_key`	`sources`, `target_entries`	`phase:standard`
`jira`	`base_url`, `email`, `api_token`	`action` (required)	`phase:standard`, `notify`
`confluence`	`base_url`, `email`, `api_token`	`action` (required)	`phase:standard`
`zendesk`	`subdomain`, `email`, `api_token`	`action` (required)	`phase:standard`
`hubspot`	`access_token`	`action` (required)	`phase:standard`
`datadog`	`api_key`, `app_key`	`action` (required)	`phase:standard`
`git`	optional `provider` (`github` \| `bitbucket`)	`action` (default `create_pr`)	`phase:standard`, `notify`
`mesh`	`url`, `client_id`, `client_secret`	`task` (required, runtime), optional `role`, `sources`	`phase:standard`

- name: fetch-docs
  agent: gdrive
  sources:
    - url: "{{inputs.gdrive_folder}}"
      name: research-sources

Gate agents

Team phases use type: gate. The phase agent is usually an llm agent (verdict in the model response). Specialized gate agents:

Agent	Settings	Inputs	Description
`hug`	`server_url` (required); optional `token`, chat/UI knobs	`role` (required), `review` (required object)	Human-in-the-loop review
`auto-gate`	—	(none)	Deterministic auto gate

LLM gates also support phase commands + max_iterations on the phase itself (orchestrator loop), not as agent manifest inputs.

Notify agents

Used for lifecycle notifications (notify capability), not as normal work phases.

Agent	Settings (required)	Inputs
`slack`	`channels` (channel → webhook map)	optional `notify` (`string[]`)
`email`	`smtp_host`, `channels` (channel → address); optional SMTP user/pass/port/tls/`from_address`	notify wiring

Meta agents

Agent	Inputs	Description
`team`	`team` (required) — e.g. `@acme/security-scan`; optional `team_inputs`	Run a nested team as one phase (`type: team` on the phase)

- name: call-universe
  type: team
  team: "@cliq/hello-universe"
  depends_on: [greet]

Choosing an agent

You need…	Prefer
Write / edit code with a coding tool	`cursor`, `claude-code`, `codex`, `gemini`
Pure LLM call without a coding CLI	`claude-api`, `gemini-api`, `openai-api`
Shell checks or scripts	`exec`
Fetch/push to SaaS or object storage	connector for that system
Human approval	`type: gate` + `agent: hug`
Nested pipeline	`type: team` + `team: @scope/slug`

List what your daemon has registered:

cliq agent list

​Agent types

​LLM agents

​Exec agents

​Connector agents

​Gate agents

​Notify agents

​Meta agents

​Choosing an agent