duvo-sandstorm 0.9.0

Open-source runtime for general-purpose AI agents in isolated sandboxes.

Sandstorm

Claude Agent SDK, or any LLM, as a sandboxed agent in your Slack, on your infra.

CLI, HTTP API, Python client, Slack bot, and TypeScript client over the same runtime. Fresh E2B sandbox per thread, streaming, file uploads, replay, and OpenTelemetry traces out of the box.

Why this exists

Anthropic shipped Claude Managed Agents (April 2026): great sandbox, Anthropic-infra only, Claude-only. Vercel shipped Slack Agent Skill + Chat SDK: great Slack wizard, Vercel-only, TypeScript-only. Neither runs on your infrastructure; neither gives you other LLMs; neither emits to your observability backend.

Sandstorm is the intersection. Self-hosted, multi-provider, Slack-native, observable. The shortest path from pip install to a production agent that runs on your terms.

Who it's for

• Teams who want Claude-style agents in Slack but can't send runtime traffic to Anthropic or Vercel.
• Engineers on GPT-5 / Gemini / DeepSeek / Qwen / Kimi / Grok who still want the Claude Agent SDK's tool + skill + sandbox story.
• Anyone running Langfuse / Phoenix / Langsmith who refuses to operate blind because their agent vendor keeps telemetry in a closed console.

What a run looks like

$ ds init research-brief && cd research-brief
$ ds "Research Acme's competitors, crawl their sites and recent news, and write a one-page briefing PDF with sources."

[tool: WebFetch]  competitor product pages + recent coverage
[tool: WebSearch] launches, pricing, positioning, buyer signals
[tool: Write]     briefing.pdf
[tool: Write]     reports/sources.md

--- Result: success | turns: 14 | cost: $0.0731 ---

artifacts:
  - briefing.pdf
  - reports/sources.md

Same prompt from Slack:

you:       @Sandstorm research Acme's competitors and post the briefing here
bot:       (paused thread sandbox resumed)
           🔧 WebFetch
           🔧 WebSearch
           🔧 Write
           Acme's top three competitors are ...
           [briefing.pdf uploaded]
           Model: sonnet | Turns: 14 | Cost: $0.0731 | Duration: 47.2s

Three things only OSS can do

1. Self-host: runs on your Railway/Fly/K8s/Docker. Data stays in your network. E2B is available self-hosted; Anthropic's Managed Agents are Anthropic-infra only.
2. Every LLM: Anthropic direct, OpenRouter (GPT-5, Gemini, DeepSeek, Qwen, Kimi, Grok), Vertex AI, Bedrock, Azure Foundry, any OpenAI-compatible base URL.
3. Your observability: OTel export to Langfuse, Phoenix/Arize, Langsmith, or any OTLP backend. No vendor console.

Sandstorm vs the closed alternatives

Feature	Sandstorm	Claude Managed Agents	Vercel Slack Agent Skill	claude-code-slack-bot
Self-hosted	✅	❌	❌	✅ (local)
Sandboxed per thread	✅	✅	✅	❌
Slack in the box	✅	❌	✅	✅
Multi-provider (not just Claude)	✅	❌	✅	❌
OTel traces to any backend	✅	❌	❌	❌
OSS license	MIT	Proprietary	Apache templates	MIT
Ships a TypeScript client	✅	✅	✅	❌
Session resume + replay	✅	✅	❌	❌

Full side-by-side in docs/comparison.md; "when should I pick each?" in docs/faq-vs-managed-agents.md.

60-second quickstart

pip install duvo-sandstorm
ds doctor                  # verify creds before running anything
ds init research-brief
cd research-brief
ds "Research Acme's competitors, crawl their sites, write a one-page brief with sources"

ds doctor is the fastest way to catch a missing ANTHROPIC_API_KEY / E2B_API_KEY / Slack scope before your first real query.

In Slack

pip install "duvo-sandstorm[slack]"
ds slack setup             # interactive, opens Slack app install in your browser
ds slack start             # Socket Mode (dev), use --http for production

Once installed:

• @Sandstorm review PR 123 in owner/repo: agent picks up context, runs tests, posts review
• /remember shipping address is Berlin: persisted per-user, per-workspace memory
• /forget shipping address: case-insensitive substring delete
• /memories: list what the bot remembers
• /model claude-haiku-4-5-20251001: per-thread model override

Thread continuity is real: each thread keeps its own paused E2B sandbox, so uploaded files, generated outputs, and installed packages survive across messages, even across server restarts. See docs/memory.md.

Pick a starter

Starter	Use when…	Aliases
general-assistant	One flexible agent for mixed workflows
research-brief	Research a topic, compare options, support a decision	competitive-analysis
document-analyst	Review transcripts, reports, PDFs, or decks
support-triage	Triage tickets into priorities, owners, next actions	issue-triage
api-extractor	Crawl docs and draft an API summary + OpenAPI spec	docs-to-openapi
security-audit	Structured security review with sub-agents and an OWASP skill
code-review	Review GitHub PRs with the GitHub MCP, inline comments, CI logs	pr-review

ds init --list for the current catalog.

Add toolpacks

ds add --list      # available: linear, notion, firecrawl, exa, github
ds add linear
ds add github

Each command wires the MCP server into sandstorm.json, seeds .env.example with the required API key, and makes the tools available to CLI / API / Slack runs from that project.

Replay a run with a different model

ds replay <run_id> --model claude-haiku-4-5-20251001 --budget 0.05

Forks the original agent session so the replay starts with identical context but on a fresh branch. A fast, cheap way to A/B compare models or reproduce bug-report runs. Reports cost Δ, latency Δ, and turn-count Δ as a markdown table. Details in docs/replay.md.

Install extras

pip install duvo-sandstorm                    # CLI + server
pip install "duvo-sandstorm[client]"          # Async Python client
pip install "duvo-sandstorm[slack]"           # Slack bot
pip install "duvo-sandstorm[telemetry]"       # OpenTelemetry

TypeScript client: npm install @duvo/sandstorm-client (see clients/typescript).

Deploy

• Railway: one-click template, 30-second deploy
• Docker: Dockerfile in repo root, docker-compose.yml for local
• Self-host: pipx install duvo-sandstorm && ds serve
• Langfuse-bundled local stack: docker compose -f deploy/docker-compose.langfuse.yml up

Docs

• Getting started
• Comparison vs Managed Agents / Victor / others
• When to pick which (FAQ)
• Multi-LLM: GPT-5, Gemini, DeepSeek, Ollama, self-hosted
• Observability: Langfuse, Phoenix, Langsmith
• Memory: /remember, /forget, persistence guarantees
• Replay
• Python client · TypeScript client
• Configuration · API reference
• Slack bot · Deployment · OpenRouter

Community

If Sandstorm closes a gap for you, star the repo and open an issue or discussion with your use case. We read every one.