Open-source infra
for Claude Agent SDK.

Deploy Claude agents to production. Your server, your sandbox, your VPC. Self-host or let us manage.

$npm install -g @ash-ai/cli
0.4ms
Overhead per message
100%
Open source runtime
$0
To self-host forever
The Shift

Developers are ditching custom harnesses for Claude Code.

The Claude Agent SDK isn't just an API client — it's a full agent runtime. Teams are ripping out LangChain, custom harnesses, and hand-rolled tool systems and replacing them with a CLAUDE.md and the SDK.

What's being replaced
LangChain agentsCustom tool registriesHand-rolled ReAct loopsJSON schema tool defsState machine orchestrators

The punchline: If Claude Code is your agent runtime, you still need production infrastructure around it. That's Ash.

File system is memory

The agent reads and writes files as working memory. Persistent, searchable, diff-able. No vector DB needed.

Skills > tool calls

Skills are markdown prompts that compose naturally. No JSON schema, no serialization overhead.

CLAUDE.md = system prompt

Plain markdown in your repo. Version-controlled with git. Reviewed in PRs. No proprietary config.

MCP for integrations

Model Context Protocol gives agents access to databases, APIs, and services through a standard interface.

Why Self-Host

Stop renting sandboxes. Own your agent platform.

Hosted platforms charge you a premium to run agents on their infrastructure. With Ash, you run everything on your own cloud—your AWS credits, your VPC, your rules.

Use your own cloud credits

Run on your existing AWS, GCP, or Azure accounts. No markup on compute. Use reserved instances, spot pricing, and existing commitments.

Data never leaves your VPC

Agent conversations, tool outputs, and session data stay in your infrastructure. Meet compliance requirements without trusting a third party.

Full source code access

Inspect, audit, extend, and modify the runtime. No black boxes. Fork it, patch it, contribute back. Open source under BSL.

MCP servers stay private

Your MCP servers run inside your VPC alongside your agents. No need to expose databases, APIs, or internal tools over the public internet.

Zero vendor lock-in

Your agents run on standard infrastructure. Switch clouds, go multi-region, or run on-prem. If we disappear tomorrow, your agents keep running.

Hosted Platforms

Third-party managed

Pay premium markup on compute
Data leaves your infrastructure
Vendor lock-in by design
Proprietary, closed-source
Subject to their outages
MCP servers exposed to the internet
Limited customization

Ash (Self-Hosted)

Your infrastructure

Use your own cloud credits & instances
Data stays in your VPC
Standard infra, no lock-in
Open source, fully auditable
Your uptime, your control
MCP servers stay private in your network
Fork, extend, customize anything
> Stop sending your agent's data to someone else's cloud. Your conversations, tool outputs, and session state belong in your infrastructure—not a third-party platform you can't audit_

Hosted agent platforms create a new dependency you don't need. You already have cloud infrastructure. You already have security policies. Ash works with what you have—not instead of it.

The Problem

You want to build this. You end up building all of this.

What you want to build
support-bot/
CLAUDE.md                  # System prompt
.claude/
  settings.json            # Tool permissions
  skills/
    search-docs/SKILL.md   # Reusable workflow
    escalate/SKILL.md      # Reusable workflow
.mcp.json                  # MCP servers
setup.sh                   # Init script
~a few files

Markdown, JSON, and bash. That's it.

What you end up building first

Session Storage

DB schema, state machine, checkpoint/restore, forking

SSE Streaming

Backpressure, reconnection, heartbeats, fan-out

Sandbox Pool

Process isolation, warm cache, idle sweep, cleanup

REST API

Routing, auth, validation, error handling, docs

Credentials

OAuth, key rotation, per-session scoping

Observability

Event logging, tracing, debugging, metrics

~100,000 lines

Before you write a single line of agent logic

Building agents is product work. Building agent infrastructure is a distraction.

Get production-ready infrastructure with Ash
Platform

Everything you need to ship agents to production.

Runtime infrastructure that handles the hard parts, plus observability to keep them running well. Self-host the runtime, optionally connect Ash Cloud for dashboards and monitoring.

Session Persistence

Database-backed session storage out of the box. Conversation history, session state, metadata—all persisted automatically with PostgreSQL or SQLite.

TypeScript SDKPostgreSQLSQLite
terminal
# Persistence is built in — just start with Postgres
$ ash start --database-url postgresql://localhost/ash
✓ Server running on :4100
✓ Connected to PostgreSQL

# Or use SQLite by default (zero config)
$ ash start
✓ Using SQLite at ~/.ash/ash.db

# Sessions, messages, metadata — all persisted
# Survives restarts automatically

Session Persistence

Database-backed session storage out of the box. Conversation history, session state, metadata—all persisted automatically with PostgreSQL or SQLite.

SSE Streaming

Real-time Server-Sent Events with backpressure, reconnection, and multi-client fan-out. No more building streaming infrastructure from scratch.

REST API Server

Auto-generated REST endpoints for sessions, messages, and agent management. Auth middleware, validation, error handling—all included.

Session Resume & Fork

Resume interrupted sessions or fork them for exploration. Branch conversation history like git branches—built into the SDK.

Run Inspector

Interactive debugger for every agent session. See the full conversation, tool calls, reasoning steps, and errors—all in one timeline.

Live Tail

Watch agent sessions in real-time as events stream in via SSE. Perfect for development and debugging active sessions.

Quick Start

Running in 5 Minutes

Install the CLI, write a CLAUDE.md system prompt, run ash start. You get a REST API, SSE streaming, session persistence, and sandboxed execution. Self-host anywhere Docker runs.

ash startash deployash session create
0.4ms

Overhead Per Message

5min

To Production

100%

Open Source Runtime

10K

Events / Minute

How It Works

Define, deploy, observe.

Write your agent logic. Ash handles the production infrastructure. Optionally connect Ash Cloud for dashboards and real-time monitoring.

Your Agent LogicYou write this
CLAUDE.md
Tools
MCP Servers
SDK import
Ash RuntimeSelf-hosted
Sessions
Streaming
REST API
Storage
Sandbox
Credentials
telemetry (optional)
Ash CloudOptional
Dashboard
Run Inspector
Live Tail
Alerts
Analytics
Team
1

Define Your Agent

Create a folder with a CLAUDE.md system prompt. Add tool permissions and MCP servers as needed. That's your entire agent.

2

Deploy on Your Infra

Run ash start anywhere — your laptop, a VM, ECS, Kubernetes. You get a REST API, SSE streaming, session persistence, and sandboxed execution. Your cloud, your credits.

3

Observe in Real-Time

Every session emits structured telemetry automatically. Debug with Run Inspector, monitor with Live Tail. Connect Ash Cloud for dashboards, or pipe events to your own stack.

Built-in telemetry — no instrumentation needed

Every session automatically emits structured events. Stream via SSE during development, or connect Ash Cloud for production monitoring.

Live Streamsession_abc123
1lifecycle{ status: "active", agent: "support-bot" }
2message{ role: "user", content: "Help me debug..." }
3reasoning{ step: "analyzing request context..." }
4tool_start{ name: "searchDocs", args: { query: "..." } }
5tool_result{ name: "searchDocs", result: { count: 5 } }
6text{ delta: "Based on the documentation..." }
7turn_complete{ tokens: 312, duration_ms: 2847 }
messagetool_starttool_resultreasoningerrorlifecycletextturn_complete
All events typed with TypeScript

Deploy Anywhere

Claude SDK
PostgreSQL
AWS / GCP / Azure
TypeScript
Docker
MCP
Performance

How much latency does Ash add? Almost none.

99.6% of latency is the Claude API — not Ash. Running on your own infra means zero network hops to a third-party platform.

0.4ms

Ash overhead per message (p50)

Target was <5ms. We hit 0.41ms.

Message Hot Path Breakdown

E2E Latency105.6ms
First Event (TTFT)54.3ms
Server Lookup0.16ms
Ash Overhead0.41ms

Session Lifecycle (SQLite)

Cold start
44ms142ms
Warm resume
1.7ms1.9ms
Cold resume
32ms36ms
p50p95

Sandbox Pool Operations

28µs
State transition
3µs
Count sandboxes
95µs
LRU eviction
265µs
Idle sweep
Pricing

Open source runtime. Free forever.

The Ash Runtime is open source and runs on your infrastructure at no cost. Ash Cloud adds dashboards, real-time monitoring, and team collaboration.

Self-Hosted

Open source, run anywhere

$0

forever

  • Unlimited agents
  • Full runtime — sessions, streaming, REST API
  • PostgreSQL or SQLite storage
  • Run Inspector + Live Tail (CLI)
  • Deploy on AWS, GCP, Azure, or bare metal
  • Full source code access
View on GitHub

Ash Cloud

Dashboards + monitoring

Free

to start

  • Everything in Self-Hosted
  • 2,500 events/month free
  • Web dashboard + Run Inspector
  • Session search and analytics
  • Team collaboration
  • Managed compute (optional)
Get Started

Enterprise

Security, compliance & scale

Custom
  • Unlimited seats + events
  • 1-year retention
  • SSO / SAML + audit logs
  • Dedicated compute pools
  • 99.9% SLA
  • Dedicated support engineer
Contact Us

Own your agent infrastructure.

Self-host in 5 minutes. Free and open source, forever.