Manus Flex lets developers bring their own inference key

Manus has launched Manus Flex, a bring-your-own-key module that lets users connect a supported inference provider and choose the model powering a Manus agent. Manus continues to run the orchestration layer that plans tasks, calls tools, manages execution environments, and turns model responses into completed work.

OpenRouter, Fireworks, and Modal are the initial inference partners. Separating model procurement from the managed agent gives teams a way to use existing provider contracts, open-weight models, and fine-tuned deployments without rebuilding their Manus workflows.

The agent stack comes apart

Manus previously bundled model selection with its agent harness and infrastructure. Flex moves model selection to the user while retaining Manus services such as browser automation, sandboxes, web app hosting, and databases.

Each Flex configuration includes a primary model and a reasoning-effort setting for the connected provider. Reasoning effort controls how much inference-time computation compatible models devote to a task, which can affect latency, cost, and output quality.

One task, two bills

A Flex task can generate charges in two accounts because inference and agent infrastructure remain separate services.

This split lets teams attribute model usage to a provider account while retaining Manus infrastructure. It also introduces separate invoices, quotas, rate limits, and service status to monitor.

Three routes to a wider catalog

Manus names OpenRouter, Fireworks, and Modal as its launch partners. The interface also displays options associated with OpenAI, Anthropic, Google AI Studio, xAI, and Meta. Those model choices can be exposed through the supported inference connections rather than requiring Manus to operate every model directly.

• OpenRouter provides access to hundreds of models through one API and billing account.
• Fireworks hosts open-weight and fine-tuned models on managed inference infrastructure.
• Modal supports custom and open-source deployments on serverless GPU infrastructure.

The partner mix covers aggregated APIs, managed open-model inference, and custom deployments. That range allows Flex to support more model configurations than the standard Manus variants alone.

Control brings new test work

Developers can route a workflow to a specific model, apply existing provider spend, and use specialized deployments while preserving the Manus planner, tools, and execution environments. Existing projects therefore avoid a harness rewrite when moving to Flex.

Changing the underlying model can still alter tool selection, structured output, context handling, latency, and refusal behavior. Teams evaluating a model should test:

• Tool and function-call accuracy
• Structured output and schema compliance
• Context-window limits and long-task performance
• Provider rate limits, timeouts, and failure handling
• Cost, latency, and completion quality on representative tasks

Model operations also move closer to the application team. Provider outages, quota changes, model deprecations, and pricing updates can affect a Flex workflow even when the Manus infrastructure remains unchanged.

Flex follows the agent market

Agent products have increasingly separated models from orchestration. Cursor and Cline support user-supplied keys, while frameworks such as LangGraph and CrewAI were designed to work across model providers. Manus approaches the same pattern from a hosted product that previously managed model selection as part of the service.

The modular design gives Manus a straightforward path to add inference partners without replacing its planning, tool-use, or execution systems. It also makes the boundary clearer for developers: providers supply model computation, while Manus supplies the agent runtime and task infrastructure.

When Flex fits

Flex suits teams with committed provider spend, negotiated rates, preferred open-weight models, or fine-tuned deployments on Fireworks or Modal. It provides model control while preserving the existing Manus workflow and infrastructure.

The managed Manus variants remain the lower-maintenance option for teams that want Manus to select models and manage inference as part of one service. Flex adds provider management, model evaluation, and a second billing stream in exchange for greater control over the model layer.