Back to blog

Claude Code Teams vs Self-Hosted: What Devs Need to Know

Adam King ·
comparison claude-code self-hosted

Looking for a Claude Code Teams alternative? Anthropic’s Claude Code Teams gives engineering organizations a managed way to deploy Claude Code across their team. It’s a good product with real strengths. But depending on your priorities around cost, data control, and multi-agent throughput, a self-hosted AI coding agent setup with a tool like Stoneforge might be a better fit.

This post compares both approaches honestly. We’ll look at what each offers, where one wins over the other, and who should pick which.

What Claude Code Teams Offers

The Claude Team plan is built for groups of 5 to 150 developers. Anthropic offers two seat types:

Standard seats ($25/month) include Claude chat, projects, and connectors with 1.25x the usage of the Pro plan. Premium seats ($150/month) add Claude Code access with 6.25x usage and early access to new collaboration features.

Administrative controls include centralized billing, SSO with domain capture, just-in-time provisioning, role-based permissions, spend caps, and Claude Code usage analytics.

Anthropic also shipped an experimental Agent Teams feature in early 2026, which lets one Claude Code session coordinate multiple “teammate” sessions for parallel work within a single codebase. It’s disabled by default and still marked experimental, but it signals where the product is headed.

What a Self-Hosted AI Coding Agent Looks Like

Self-hosting means running Claude Code (or any compatible agent) against your own API keys, with an orchestration layer managing the workflow. With Stoneforge, the stack looks like this:

  • Your API keys with direct Anthropic API access (or any other provider)
  • Agents running on your machines or CI/CD runners
  • Code that never leaves your network except for model API calls
  • An orchestration daemon that manages parallel agent sessions, task dispatch, dependency tracking, and merge coordination
# Self-hosted setup with Stoneforge
sf init
sf config set provider anthropic
sf config set api-key $ANTHROPIC_API_KEY
sf daemon start --workers 3

Stoneforge is open-source under Apache 2.0. There’s no license fee, no seat count, and no vendor lock-in on the orchestration layer.

Cost Comparison

Pricing changes over time. These numbers are current as of March 2026. Check Anthropic’s pricing page for the latest.

Claude Code Teams

Premium seats (required for Claude Code access) cost $150/month per developer. That includes a generous usage allocation, but overages are billed per-token. For a team of 10 developers, the base cost is $1,500/month before any overages.

Standard seats at $25/month don’t include Claude Code, so they’re not relevant to this comparison.

Self-Hosted (API Direct)

You pay Anthropic’s API pricing directly. No seat fees. With Claude Sonnet, that’s $3 per million input tokens and $15 per million output tokens.

The Math

For a developer running roughly 50 agent tasks per week, with an average of 20k input tokens and 4k output tokens per task:

  • Input: 50 tasks x 20k tokens x 4 weeks = 4M input tokens = $12/month
  • Output: 50 tasks x 4k tokens x 4 weeks = 800k output tokens = $12/month
  • Total: ~$24/developer/month at API list price
  • With prompt caching (typical 60-70% cache hit rate): ~$12-15/developer/month

That’s roughly 6-12x cheaper than a $150 premium seat. For a team of 10, the difference is $1,500/month vs. $120-240/month.

The managed offering bundles platform support, admin tooling, and the convenience of not managing API keys. Whether that premium is worth it depends on your team.

Caveat: These numbers assume moderate usage. Developers who run agents continuously will see higher API bills, and the gap narrows. The breakeven depends on your specific usage patterns.

Data Privacy and Security

This is often the deciding factor for enterprise teams.

Claude Code Teams

Code context is sent to Anthropic’s infrastructure for processing. Team and Enterprise plans don’t use your data for model training by default. Enterprise customers can negotiate a Zero-Data-Retention (ZDR) addendum for maximum data isolation, and custom retention controls allow a minimum 30-day retention period.

Anthropic does not currently offer regional data residency guarantees, which matters for organizations operating under GDPR or similar frameworks.

Self-Hosted

Code stays on your machines. Only model API calls leave your network, and you control exactly what those calls contain. You can run behind a VPN, route through an API proxy for auditing, or use a self-hosted model for complete air-gapped operation.

The audit trail and data lifecycle are entirely under your control, which simplifies compliance with SOC 2, HIPAA, FedRAMP, and similar frameworks.

For teams working on proprietary algorithms, financial systems, healthcare data, or government contracts, self-hosting eliminates an entire category of compliance conversations.

Multi-Agent Capabilities

Both approaches now support multi-agent work, but with different architectures.

Claude Code Agent Teams

Anthropic’s experimental Agent Teams feature lets one Claude Code session act as a team lead, coordinating teammate sessions that work in parallel. Teammates can share context, communicate directly, and converge on solutions together.

The strongest use cases are research (multiple teammates investigating different aspects simultaneously), building new modules (each teammate owns a separate piece), and debugging with competing hypotheses.

It’s a meaningful step forward. Two limitations to note: it’s still experimental (disabled by default, enabled via CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS), and it operates within a single developer’s session rather than across a team.

Self-Hosted with Stoneforge

Stoneforge takes a different approach. Instead of teammates within one session, it manages fully independent agent processes across a team, each in its own isolated git worktree:

  • A daemon dispatches tasks from your backlog to available worker agents
  • Each worker runs in complete isolation with no shared context to corrupt
  • A merge steward handles branch integration, conflict resolution, and CI checks
  • Agents are assigned specialized roles: Director plans work, Workers execute, Stewards review
  • The system tracks dependencies so blocked tasks wait automatically
# Stoneforge: daemon manages the full lifecycle
$ sf daemon start --workers 5
# 5 independent agents pick tasks, work in parallel,
# push branches, and merge steward integrates results

The architectural difference: Agent Teams coordinates within a session for a single developer. Stoneforge coordinates across sessions for an entire team, with persistent task tracking, automatic recovery when agents fail, and a merge pipeline that keeps main green.

Flexibility and Customization

Claude Code Teams is Claude-only. Anthropic controls the feature roadmap, and configuration is limited to what the platform exposes.

Self-hosted orchestration is agent-agnostic. Stoneforge works with Claude Code, Codex, OpenCode, or a mix. You can customize agent prompts and roles per task type, integrate with your existing CI/CD and issue trackers (Linear sync, GitHub issues), and build custom stewards for team-specific review workflows. It’s open-source, so you can inspect and modify anything.

When to Choose Claude Code Teams

Claude Code Teams is the right choice when:

  • You want zero operational overhead with no infrastructure to manage
  • Your team is small (under 10 developers) and $150/seat/month is acceptable
  • Your code isn’t subject to strict data residency or compliance requirements
  • You primarily need single-developer AI assistance rather than team-wide orchestration
  • You value official Anthropic support and a managed experience

When to Choose a Self-Hosted AI Coding Agent

Self-hosting with Stoneforge as a Claude Code Teams alternative makes more sense when:

  • Cost matters at scale. API-direct pricing is substantially cheaper, especially for teams above 5 developers.
  • Data must stay on your infrastructure for compliance, IP protection, or policy reasons
  • You need team-wide orchestration with persistent task queues, automatic dispatch, and merge management
  • You want provider flexibility to switch between or combine different AI coding agents
  • Customization is important: control over agent behavior, prompts, review workflows, and integrations
  • You’re already running infrastructure and adding Stoneforge is low marginal effort

For a detailed feature-by-feature breakdown, see our Stoneforge vs Claude Code Agent Teams comparison.

Making the Decision

Claude Code Teams is a well-engineered managed service that’s getting better quickly. The Agent Teams feature, once it graduates from experimental status, will close the multi-agent gap for single-developer workflows.

For individual developers or small teams that want AI coding assistance without thinking about infrastructure, Claude Code Teams is the pragmatic choice.

For teams that care about cost optimization, data sovereignty, team-wide orchestration, or workflow customization, self-hosting is worth the setup investment. Stoneforge’s getting started guide takes about 15 minutes.

Stoneforge is a young open-source project. It won’t have the polish of a managed Anthropic product, and you’ll be responsible for your own infrastructure. That trade-off is the core of this decision: managed convenience vs. control and flexibility. Both are valid.

Frequently Asked Questions

Can I migrate from Claude Code Teams to self-hosted?

Yes. Both approaches use Claude Code as the underlying agent. Your .claude/ configuration, CLAUDE.md files, and workflow knowledge transfer directly. Stoneforge adds orchestration on top without modifying how Claude Code itself operates.

Does self-hosting mean I lose Anthropic support?

You lose platform-level support (the Teams admin features), but you still use Anthropic’s API with their standard API support. Stoneforge has its own community and documentation. For most teams, the Claude Code CLI docs plus Stoneforge’s orchestration docs cover what you need.

What about latency?

Identical. In both cases, the agent makes API calls to Anthropic’s model serving infrastructure. The model inference happens on the same servers regardless of whether you’re on a Team plan or using API keys directly.

How does Claude Code Agent Teams compare to Stoneforge’s multi-agent system?

Agent Teams coordinates multiple Claude Code sessions within a single developer’s workflow, with a team lead assigning tasks to teammates. Stoneforge orchestrates fully independent agent processes across an entire team, with persistent task tracking, automatic dispatch from a shared backlog, and a merge pipeline. They solve related but different problems: Agent Teams optimizes individual developer throughput, Stoneforge optimizes team-wide throughput.

Can I use both together?

Yes. You could use Stoneforge for team-wide orchestration while individual worker agents use Agent Teams internally for complex subtasks. The two layers don’t conflict.