What you will have at the end
A local Promptsheon instance with the Fastify API, Next.js console, SQLite persistence, and a first agent specification ready to execute and evaluate. The default setup is intentionally local and self-hosted so you can inspect every boundary.
1. Install and start
git clone https://github.com/sachncs/promptsheon.git
cd promptsheon
pnpm install
cp .env.example .env
pnpm dev Open http://localhost:3000. The API runs on http://localhost:8080. Onboarding creates the first administrator, organization, and provider configuration.
2. Configure a provider
Choose OpenAI, Anthropic, Bedrock, or an OpenAI/Anthropic-compatible custom endpoint. Credentials are stored through the vault boundary and are never written to system_config in plain text. For an unattended environment, provide the provider credentials through the deployment secret manager before boot.
3. Create an AgentSpec
An AgentSpec is an immutable composition of prompts, nodes, tools, policies, runtime limits, and evaluation intent. Start small: one classification step, one tool or retrieval step, and one response step with a guardrail.
name: refund-triage
version: 1
graph:
- classify: { model: claude-3-5 }
- retrieve: { tool: orders }
- respond: { guardrail: pii-redact } 4. Observe and evaluate
Attach a dataset and run deterministic or LLM-judge scorers. Review outcome quality alongside latency, token use, cost, tool calls, and failures. The evidence should explain not only whether the run passed, but what to improve next.
5. Validate and promote
Compare a candidate against the last known-good content hash through replay or controlled traffic. Promote only when quality and efficiency improve under the same workload; keep the previous identity available for recovery.
Where to go next
- Set up local development for focused server, frontend, or site work.
- Run the test tiers before sharing a change.
- Read the architecture guide to understand package and dependency boundaries.
- Integrate the API and SDK into CI or your own runtime.
- Prepare a production deployment with health checks and backups.