AI plugin reference

These processors call a chat model, call tools on another MCP server and keep conversation history. LLM_CALL is in module llm-call, MCP_LIST_TOOLS and MCP_TOOL_CALL in mcp-tool-call, SESSION_MEMORY in session-memory. All option values support ${...} unless a row says otherwise. Write secrets as decrypted:<key>.

`LLM_CALL`

Calls a chat model and emits its answer. The node is blocking.

Option	Type	Default	Required	Description
`provider`	`openai`, `openai-compatible`, `anthropic` or `ollama`	none	yes	Provider; `openai-compatible` is the same as `openai`.
`model`	string	none	yes	Model name.
`prompt`	expression	none	`prompt` or `messages`	Appended as the last user turn.
`messages`	list or expression	none	`prompt` or `messages`	Turns `[{role, content}]` with roles `user`, `assistant`, `tool`, or a SpEL expression such as `payload['history']`; `${...}` turns a list into text and is not for this option.
`system`	expression	none	no	System prompt.
`baseUrl`	string	provider default	no	Endpoint override (Groq, vLLM, NIM, Ollama).
`apiKey`	secret	none	no	API key.
`temperature`	number	model default	no	Sampling temperature.
`maxTokens`	number	model default	no	Most output tokens.
`timeout`	number	`120`	no	Seconds per attempt, positive.
`retries`	number	`0`	no	Extra attempts with exponential backoff starting at 500 ms; not negative.
`tools`	list or expression	none	no	`[{name, description, inputSchema}]`, for example the `tools` of an `MCP_LIST_TOOLS` node. Calls come back in `actions` and are not run by this node.
`allowAskInput`	boolean	`false`	no	Offer the model the tool `ask_for_input {prompt, options?, kind?, risk?}`; wire its action to an `MCP_TOOL_CALL` node with `allowAskInput`.
`prices`	map	none	no	`{input, output}` price per million tokens; adds `cost`. Accepted, not in the declared schema.
`budget`	map	none	no	`{maxTokens, maxCost, maxCalls, agent, hardStop}` for all calls of one run; `maxCost` needs `prices`.
`failureMode`	`ERROR` or `CONTINUE`	`ERROR`	no	`CONTINUE` emits the original payload with metadata `llmSuccess=false` and `llmError`.
`routeByStatus`	boolean	`false`	no	Emit on `success` or `failure`; needs `failureMode: CONTINUE`.
`concurrency`	number	`16`	no	Model calls in flight at once.

Output ports: default, or success and failure with routeByStatus.
Payload: {success, content, actions, usage{inputTokens, outputTokens}, latencyMs, model, cost?}. actions holds {tool, arguments, id} entries, also parsed from <agent_action> blocks. Metadata gains llmSuccess=true.
Budget: the ledger is shared by every LLM_CALL and MCP_TOOL_CALL node of the run (a budget with agent counts that agent apart). The check runs before a call and usage is added after it, so one call can overshoot. When exhausted, the next call fails like a failed call and failureMode applies; with hardStop: true it ends the whole run instead, and nothing retries, drops or routes it.
Default node timeout: (timeout + 1) * (retries + 1) + 5 seconds plus the backoff.
Server settings: yukta.llm.deny-input-patterns, yukta.llm.deny-output-patterns (lists of regular expressions; a match refuses the call) and yukta.llm.rate-limit.<provider> (requests per minute). See Configuration.
Errors: messages or prompt is mandatory; messages and/or prompt must yield at least one turn; timeout must be positive, got: <n>; retries must not be negative; allowAskInput must be true or false; routeByStatus requires failureMode: CONTINUE; LLM call failed: <cause>.

{
  "nodeId": "ask-model",
  "type": "LLM_CALL",
  "config": { "provider": "anthropic", "model": "claude-sonnet-5-5", "apiKey": "decrypted:anthropic", "prompt": "${payload.message}", "budget": { "maxTokens": 50000 } }
}

`MCP_LIST_TOOLS`

Lists the tools of a remote MCP server (streamable HTTP) so a model can choose among them.

Option	Type	Default	Required	Description
`url`	string	none	yes	MCP endpoint, `http` or `https`.
`headers`	map	none	no	Extra request headers.
`authToken`	secret	none	no	Bearer token.
`timeout`	number	`30`	no	Seconds, positive.
`allowedTools`	list	none	no	Keep only these tools.

Output port: default. Payload: {tools: [{name, description, inputSchema}], count}, replacing the incoming payload.
Paging: the node follows nextCursor for up to 50 pages and fails if the list does not end by then.
Default node timeout: timeout plus 5 seconds.
Errors: url is mandatory; url must be http(s), got: <url>; timeout must be positive, got: <n>.

{ "nodeId": "tools", "type": "MCP_LIST_TOOLS", "config": { "url": "https://mcp.example.com/mcp", "authToken": "decrypted:mcp" } }

`MCP_TOOL_CALL`

Calls a tool on a remote MCP server per message.

Option	Type	Default	Required	Description
`url`	string	none	yes	MCP endpoint, `http` or `https`.
`tool`	string	none	yes	Tool name.
`arguments`	expression	the payload if a map, else `{"input": payload}`	no	A map of values, or a SpEL string returning a map.
`headers`	map	none	no	Extra request headers.
`authToken`	secret	none	no	Bearer token.
`timeout`	number	`60`	no	Seconds, positive.
`budget`	map	none	no	`{maxCalls, agent, hardStop}`, counted with the `LLM_CALL` budget of the same run.
`allowedTools`	list	none	no	Tools this node may call; any other tool, even one chosen through `${...}`, is refused.
`requireApproval`	list	none	no	Tools (or `*`) that wait for a person before they are called.
`approvalSignal`	string	`approve`	no	Name of the approval signal.
`approvalTimeoutSeconds`	number	`3600`	no	Longest wait; no answer refuses the call.
`allowAskInput`	boolean	`false`	no	Serve the tool `ask_for_input` as an input request.
`inputRiskFloor`	string	`MEDIUM`	no	Lowest risk of a question from the model; it cannot ask for less.
`inputTtlSeconds`	number	`3600`	no	Seconds a question from the model lives.
`inputAudience`	list	current tier	no	Who may answer; the model cannot set it.
`inputRoute`	string	none	no	Route of the questions.
`inputMaxAsks`	number	`3`	no	Most questions per node run.
`inputOnTimeout`	`ERROR` or `FAIL`	`ERROR`	no	No answer in time: a tool error the model sees, or a failed node.
`resultKey`	string	none	no	Also copy the text result into this metadata key.
`failureMode`	`ERROR` or `CONTINUE`	`ERROR`	no	`CONTINUE` emits the result with `success=false`.
`routeByStatus`	boolean	`false`	no	Emit on `success` or `failure`; needs `failureMode: CONTINUE`; edges may then use only `success`, `failure` and `error`.
`concurrency`	number	`16`	no	Tool calls in flight at once.

Output ports: default, or success and failure with routeByStatus.
Payload: {success, tool, text, content, structured, error}. Metadata gains mcpSuccess, mcpTool, mcpLatencyMs.
Approval: the wait is an input request of kind APPROVAL and risk HIGH, answered by a person; the decision reject refuses the call, any other approves it.
Control: a control-bus payload reset drops the cached MCP sessions.
Default node timeout: timeout plus 5 seconds, plus approvalTimeoutSeconds when requireApproval is set, plus the ask-input wait when allowAskInput is set.
Server settings: yukta.mcp-client.max-argument-bytes (default 65536) and yukta.mcp-client.deny-argument-patterns (regular expressions). The URL is checked against yukta.egress.*.
Errors: url and tool are mandatory; url must be http(s), got: <url>; MCP tool '<tool>' is not in allowedTools [...]; arguments must evaluate to a map; allowedTools must be a list of tool names; MCP tool <tool> failed: <error>; routeByStatus requires failureMode: CONTINUE.

{
  "nodeId": "act",
  "type": "MCP_TOOL_CALL",
  "config": { "url": "https://mcp.example.com/mcp", "tool": "create_issue", "arguments": "payload.toolCall.args", "requireApproval": ["create_issue"] }
}

`SESSION_MEMORY`

Chat memory per conversation id and workflow. Needs the control plane. History is per workflow: another workflow that knows the id cannot read it.

Option	Type	Default	Required	Description
`mode`	`load`, `append` or `delete`	none	yes	`load` adds history; `append` records the user turn and the reply together; `delete` removes the history.
`sessionId`	string	`${metadata.sessionId}` for `append`	`load` and `delete`	Conversation id.
`message`	expression	none	`load`	The user turn to record.
`assistant`	expression	`${payload.content}`	no	The reply to record (`append`).
`maxTurns`	number	`20`	no	Turns kept per session and returned by `load`; positive.

Output port: default. load outputs the payload (an object; other payloads become {input: payload}) plus history: [{role, content}], and sets metadata sessionId. append and delete pass the message through.
Storage: in memory by default and lost on restart; yukta.session-memory.max-sessions (default 10000, least recently used leave first) and yukta.session-memory.ttl apply. Storing state on disk (yukta.session-memory.dir) belongs to the Plus edition; see Licensing.
Errors: maxTurns must be positive, got: <n>.

{ "nodeId": "memory-load", "type": "SESSION_MEMORY", "config": { "mode": "load", "sessionId": "${payload.sessionId}", "message": "${payload.message}" } }

Workflows

Plugins

Supervisors

Interfaces

Server

AI plugin reference

AI plugin reference

`LLM_CALL`

`MCP_LIST_TOOLS`

`MCP_TOOL_CALL`

`SESSION_MEMORY`

AI plugin reference

LLM_CALL

MCP_LIST_TOOLS

MCP_TOOL_CALL

SESSION_MEMORY

Related

`LLM_CALL`

`MCP_LIST_TOOLS`

`MCP_TOOL_CALL`

`SESSION_MEMORY`