Shape how your AI thinks

The problem

Agent config files grow into unmanageable walls of text. One monolithic prompt for every task. Copy-paste between projects. No way to switch behavior without rewriting everything. See Why Composable Prompts for the full case.

How it works

Instead of one monolithic config file, brainjar separates agent behavior into composable layers. Each layer is a markdown document under your local SQLite store (or a brainjar serve instance). Mix and match them per project or per task.

Soul Who the agent is — voice, character, standards. The constant across all tasks.

Persona How the agent works — role and disposition. Switch per task.

Rules What the agent must follow — behavioral constraints and guardrails.

Procedure The step-by-step playbook the agent runs. Swap method without changing role.

Skill On-demand capabilities loaded by the platform — runbooks, decision trees, glossaries.

Brain A saved bundle of soul + persona + rules + procedure + skills. One command to load.

Quickstart

curl -fsSL https://get.brainjar.sh/brainjar/install.sh | sh
brainjar init

That gives you a workspace, a default API key, and everything you need to start composing. Already have a large config file? Migrate it.

Built for agents

Every brainjar command takes --json for machine-readable output on stdout and a structured error envelope on stderr. The same binary doubles as an MCP server (brainjar mcp) so agents call composition, state, and content tools natively rather than shelling out.

brainjar status                # human: formatted with context
brainjar status --json         # agent: structured data

These docs are agent-readable too. Point your agent at brainjar.sh/llms.txt for a doc index, or append .md to any page URL for raw markdown (e.g. /getting-started.md).