Character

A character is the piece of the system that turns all of the underlying context (roles, assets, MCP tools, agents) into an actual copilot that a human experiences. It's where we define how it behaves and where its inferences are run.

In the architecture, characters are one of the Core Objects: they hold the inference endpoint, inference configuration, and system prompt chunks. The LLM assumes a character, like an actor taking on a role in a specific space, so that each user gets a personality and focus matching their role rather than a generic model.

The Bifrost admin tool is where characters are created, edited, and inspected. Each character appears as a configurable unit that can be safely wired into collections.

Character Config Model

The CharacterConfig schema defines how a character is represented in configuration, APIs, and in the Bifrost admin tool:

CharacterConfig {
  character_id: string
  name?: string
  description?: string
  inference_endpoint?: string
  custom_inference_uri?: string
  auth_type?: string
  auth_uri?: string
  pre_tool_prompt?: string
  post_tool_prompt?: string
  welcome_prompt?: string
  config?: object
  created_at?: number
  updated_at?: number
  customer_id?: string
}

System

Inference and Auth

Clearly define where this character's inference runs and how it proves it's allowed to run there, so that authorization and routing breadcrumbs are built into the structure, not bolted on later.

Persona (pre-tool) System Prompt

Define the character's identity and priorities before any tools or external context are added.

The Persona is encoded primarily in:

• pre_tool_prompt?: string

Think of pre_tool_prompt as the first half of the system prompt that sets the stage:

• Who the character is (role, domain expertise, vantage point).
• How it should speak to this organization's users (tone, level of detail).

Characters are meant to mesh with the user's role and environment - operators vs. technicians vs. foremen all see different behavior, even if underlying tools are similar.

Procedure (post-tool) System Prompt

Define how the character uses tools and data to execute procedures - the what you do with information once you have it.

The Procedure portion is encoded in:

• post_tool_prompt?: string

The post_tool_prompt is typically the second half of the system prompt. It says what and how to use the available tools to accomplish common tasks. This is where core procedures and workflows are embedded into the character.

Typical responsibilities for post_tool_prompt:

• Interpreting tool results
  • How to reconcile conflicting data from multiple MCP tools or agents.
• Following domain procedures
  • How to turn raw readings, events, or records into stepwise guidance that matches the organization's SOPs.
• Escalation and handoff
  • When to stop, summarize, and hand off to a human (e.g., safety-critical thresholds, ambiguous diagnosis-like situations).

Config

Tune how the character behaves at run-time: model settings, safety levers, and feature flags that adapt the character to each organization and use case. These are typically specific to the inference endpoint as each model and framework has a different set of levers to adjust tuning.

A provider- and deployment-specific JSON object used to parameterize inference and behavior. This is intentionally unopinionated so it can adapt to different backends while staying governed at the platform level.

Common uses include:

• Model & decoding parameters
  • Model name or family, temperature, max tokens, and other sampling configuration.
• Tooling behavior
  • Limits on number of tool calls per turn, timeout settings, retry strategies.

Characters Configuration In Bifrost

When you create or edit a character in Bifrost, you are:

1. Choosing where the inference runs and how it authenticates (endpoint URI & auth).
2. Defining who it is to your users (Persona / pre-tool system prompt).
3. Encoding how it executes real-world procedures using assembled tools and data (Procedure / post-tool system prompt).
4. Tuning how it behaves operationally (config).

Together, these make each character a precise, governed expression of "intelligence that understands the moment and turns awareness into action - safely" for a specific role and space in your organization.

Bridge API Endpoints