PromptRails
WE BUILD THE RAILS · YOU SHIP THE AGENTSOPEN SOURCE → PRODUCTION · 2026

Ship the
agent. Skip
the plumbing.

Promptrails is the control plane for AI agents in production —prompts, orchestration, tracing, guardrails, cost, across 14 LLM providers. One contract. No lock-in.

Start building →$pip install promptrails
What you replace
Prompt registryPromptLayer · Pezzo
Gateway / routerOpenRouter · Portkey
TracingLangfuse · LangSmith
GuardrailsGuardrails AI · NeMo
EvalsBraintrust · Patronus
Approval queue(homegrown)
01 / AI ENGINEERS

Build agents, not plumbing.

One SDK, 14 providers, versioned prompts, full tracing — Python, JS, or Go.

Read the docs →
02 / PRODUCT TEAMS

Iterate on prompts, no deploy.

Promote, A/B, roll back from the dashboard. Evals + scores in the same loop.

See versioning →
03 / ENG LEADERS

Cost & latency, by span.

Per-agent rollups, p99 dashboards, scanner verdicts. Decide what ships, with data.

View metrics →
04 / ENTERPRISES

SSO. Audit. BYO cloud.

Roles, audit log, encrypted credentials, DPA on request. SOC 2 in progress.

Talk to us →
03 / Platform

Six tools.
One contract.

Most teams ship an agent by gluing six SaaS tools together. We replaced the glue with a single SDK and a single trace.

/ 01

Agents.

Five execution strategies — simple, chain, multi-agent, workflow, composite — composed from versioned prompts and tools.

Agent.execute()
/ 02

Prompts.

Jinja2 templates with input/output schemas. Promote, roll back, A/B without a deploy.

Prompt.latest("classify")
/ 03

Tracing.

OTel-style distributed traces with 14 span kinds. Every LLM call, tool, and guardrail captured.

span.kind = "llm.chat"
/ 04

Guardrails.

14 scanners. PII, toxicity, prompt injection, jailbreak. Block, redact, or log.

guardrails: ["pii"]
/ 05

Cost.

Per-execution and per-span cost across providers. Workspace and per-agent rollups.

cost_usd: 0.0042
/ 06

MCP tools.

First-class Model Context Protocol. Connect APIs, data sources, remote MCP servers.

tools: [mcp.server("sql")]
14
LLM PROVIDERS
— ONE API
70+
INTEGRATIONS
— DBs, vector, SaaS
p99 < 18ms
GATEWAY OVERHEAD
— measured weekly
38%
AVG COST CUT
— from routing alone
04 / Topology

One graph.
Every span.

Every run is a tree of spans — calls, tools, guardrails, retrievals — with cost, latency, and the full prompt + response captured at each node.

01Drill into any span — input, output, model, tokens
02Compare prompt versions head-to-head
03Filter by cost, latency, scanner verdict, score
04Export to your existing OTel collector
EXEC GRAPH · CHAIN STRATEGY
INPUTAGENTCLASSIFYgpt-4.1TOOL · sqlRESOLVEsonnetGUARD · piiGUARD · toxOUTPUTscored · 0.0042
05 / SDK

Wherever
you write code.

Python
$ pip install promptrails
JavaScript
$ npm i @promptrails/sdk
Go
$ go get promptrails.ai/sdk
CLI
$ brew install promptrails/tap/cli
MCP server
$ claude mcp add promptrails
VSCode ext
$ code --install-extension promptrails
06 / DASHBOARD · MULTI-PAGE · PIN-PROTECTABLE · BACKED BY YOUR AGENTS
app.promptrails.ai
ProductionAgents

Agents

3 agents

+ New Agent
NameTypeStatusDescription
Customer Support Bot
multi_agent
active
Handles customer inquiries with tool access
Data Analyst
chain
active
SQL generation and analysis pipeline
Doc Summarizer
simple
archived
Summarizes documents with key points
07 / Field reports
We deleted four internal services the day we shipped Promptrails. Tracing alone paid for it inside a month — turns out half our prompts were calling GPT-5 when Haiku was fine.
Cardash
Kenneth
09 / Get on it

5 min
from pip install promptrails
to traced agent in production.

Start building →Read the docs