TokenLimiterProcessor

The TokenLimiterProcessor limits the number of tokens in messages. Depending on trimMode, it acts as a prompt processor, an input processor, and an output processor:

• Prompt processor (processLLMRequest): In the default best-fit and contiguous trim modes, enforces the input budget on the provider prompt right before each model call, at every step of the agentic loop. The prompt is measured after earlier prompt processors (such as ToolCallFilter) have transformed it, so only tokens that actually reach the model are counted. Tool call and tool result messages are grouped so they're kept or removed together, and trimming is transient: stored messages are never modified.
• Input processor (processInput): In memory-only trim mode, filters historical messages to fit within the context window before the agentic loop starts, prioritizing recent messages
• Output processor: Limits generated response tokens via streaming or non-streaming with configurable strategies for handling exceeded limits

Usage example

import { TokenLimiterProcessor } from '@mastra/core/processors'

const processor = new TokenLimiterProcessor({
  limit: 1000,
  strategy: 'truncate',
  countMode: 'cumulative',
})

Constructor parameters

options:

number | Options

Either a simple number for token limit, or configuration options object

number | Options

limit:

number

Maximum number of tokens to allow in the response

encoding?:

TiktokenBPE

Optional encoding to use. Defaults to o200k_base which is used by gpt-5.1

strategy?:

'truncate' | 'abort'

Strategy when token limit is reached: 'truncate' stops emitting chunks, 'abort' calls abort() to stop the stream

countMode?:

'cumulative' | 'part'

Whether to count tokens from the beginning of the stream or just the current part: 'cumulative' counts all tokens from start, 'part' only counts tokens in current part

trimMode?:

'best-fit' | 'contiguous' | 'memory-only'

Controls how the token limit is enforced: 'best-fit' trims the provider prompt while keeping as many messages as possible (may create gaps), 'contiguous' trims the provider prompt but stops at the first message that does not fit (keeping a continuous suffix of conversation history), and 'memory-only' trims stored history in processInput instead of the prompt

Returns

id:

string

Processor identifier set to 'token-limiter'

name?:

string

Optional processor display name

processInput?:

(args: { messages: MastraDBMessage[]; abort: (reason?: string) => never }) => Promise<MastraDBMessage[]>

Trims stored history to fit within the token limit before the agentic loop starts in 'memory-only' trim mode, prioritizing recent messages while preserving system messages and the current turn

processInputStep?:

(args: ProcessInputStepArgs) => Promise<void>

In 'memory-only' trim mode, applies stored-history trimming at each step. In the 'best-fit' and 'contiguous' trim modes, it leaves trimming to processLLMRequest when the agent runs that method for this processor. Otherwise it trims stored history, for example in generateLegacy() and streamLegacy() or when the limiter is inside a processor workflow.

processLLMRequest?:

(args: ProcessLLMRequestArgs) => Promise<ProcessLLMRequestResult>

Enforces the input budget on the provider prompt in the 'best-fit' and 'contiguous' trim modes. Runs after earlier prompt processors have transformed the prompt, counts that exact prompt, and returns a trimmed copy for the model call only. System messages are always preserved, and tool call and tool result messages are grouped so they are kept or removed together.

processOutputStream:

(args: ProcessOutputStreamArgs) => Promise<ChunkType | null>

Processes streaming output parts to limit token count during streaming. Only text and object parts count against the limit and can be withheld; lifecycle, reasoning and tool parts always pass through.

processOutputResult:

(args: { messages: MastraDBMessage[]; abort: (reason?: string) => never }) => Promise<MastraDBMessage[]>

Processes final output results to limit token count in non-streaming scenarios

getMaxTokens:

() => number

Get the maximum token limit

Output stream behavior

As an output processor, only parts that carry generated output count against the limit: text-delta and object. Lifecycle parts (such as step-start), reasoning deltas, response metadata, and tool parts (tool-call, tool-result) are neither counted nor withheld, so tool calls always reach the agentic loop and get executed.

With the default truncate strategy, the first time output is withheld the processor emits a transient data-token-limit-reached part on the stream:

for await (const part of stream.fullStream) {
  if (part.type === 'data-token-limit-reached') {
    console.log('output truncated at', part.data.limit, 'tokens')
  }
}

Media token counting

Images and file attachments are estimated instead of tokenized, including file message parts and tool results shaped like { data, mediaType }. Images use a flat per-image estimate; other media uses decoded byte size, while remote URLs and provider file ids use a flat fallback because their size isn't known locally. Encoded payloads such as base64 data are never counted as text, avoiding an inflated count that could truncate history unnecessarily.

Error behavior

When trimming input, TokenLimiterProcessor throws a TripWire error in the following cases:

• Empty messages: If there are no non-system messages to process, a TripWire is thrown because you can't send an LLM request with no messages.
• System messages exceed limit: If system messages alone exceed the token limit, a TripWire is thrown because you can't send an LLM request with only system messages and no user/assistant messages.
• No messages fit: If no message fits within the remaining token budget, a TripWire is thrown because you can't send an LLM request with no messages.

import { TripWire } from '@mastra/core/agent'

try {
  await agent.generate('Hello')
} catch (error) {
  if (error instanceof TripWire) {
    console.log('Token limit error:', error.message)
  }
}

Extended usage example

As an input processor (limit context window)

Use inputProcessors to limit historical messages sent to the model, which helps stay within context window limits:

src/mastra/agents/context-limited-agent.ts

import { Agent } from '@mastra/core/agent'
import { Memory } from '@mastra/memory'
import { TokenLimiterProcessor } from '@mastra/core/processors'

export const agent = new Agent({
  id: 'context-limited-agent',
  name: 'context-limited-agent',
  instructions: 'You are a helpful assistant',
  model: 'openai/gpt-5.6-sol',
  memory: new Memory({/* ... */}),
  inputProcessors: [
    new TokenLimiterProcessor({ limit: 4000 }), // Limits historical messages to ~4000 tokens
  ],
})

As a per-step processor (limit multi-step token growth)

When an agent uses tools across multiple steps (e.g. maxSteps > 1), each step accumulates conversation history from all previous steps. TokenLimiterProcessor applies its limit at every step, measuring the prompt that's about to be sent after any earlier prompt processors have run. Register prompt-shrinking processors such as ToolCallFilter before it, so the limiter counts the prompt the model actually receives:

src/mastra/agents/multi-step-agent.ts

import { Agent } from '@mastra/core/agent'
import { TokenLimiterProcessor } from '@mastra/core/processors'

export const agent = new Agent({
  id: 'multi-step-agent',
  name: 'multi-step-agent',
  instructions: 'You are a helpful research assistant with access to tools',
  model: 'openai/gpt-5.6-sol',
  inputProcessors: [
    new TokenLimiterProcessor({ limit: 8000 }), // Applied at every step
  ],
})

// Each tool call step will be limited to ~8000 input tokens
const result = await agent.generate('Research this topic using your tools', {
  maxSteps: 10,
})

Processor workflows don't run processLLMRequest, so a TokenLimiterProcessor inside a processor workflow trims stored messages in processInputStep, before prompt processors such as ToolCallFilter remove anything from the request. To count the prompt the model receives, add the limiter directly to inputProcessors.

As an output processor (limit response length)

Use outputProcessors to limit the length of generated responses:

src/mastra/agents/response-limited-agent.ts

import { Agent } from '@mastra/core/agent'
import { TokenLimiterProcessor } from '@mastra/core/processors'

export const agent = new Agent({
  id: 'response-limited-agent',
  name: 'response-limited-agent',
  instructions: 'You are a helpful assistant',
  model: 'openai/gpt-5.6-sol',
  outputProcessors: [
    new TokenLimiterProcessor({
      limit: 1000,
      strategy: 'truncate',
      countMode: 'cumulative',
    }),
  ],
})

Related

• Guardrails