🧭 Anthropic Launches $150M Claude Corps Fellowship to Embed AI in US Nonprofits

Anthropic has committed $150 million to launch Claude Corps, a national fellowship that will place 1,000 early-career professionals inside U.S. nonprofits for 12-month paid placements starting October 2026. Led by President Daniela Amodei, the programme is structured around three partners: Anthropic funds and designs the curriculum, CodePath recruits, trains, and employs the fellows, and Social Finance handles measurement and administers $10,000 host grants. Announced publicly on June 11 with broad press coverage today, Claude Corps represents Anthropic's most tangible acknowledgement yet that AI-driven labour displacement is real — and that it has a direct responsibility to respond.

Who can apply and what fellows do

• Eligibility: Anyone over 18 with fewer than two years of full-time work experience, authorised to work in the US, comfortable using Claude, and willing to relocate to a host city if required. No degree requirement.
• Compensation: $85,000 annual salary plus benefits, paid by CodePath (not the host nonprofit).
• Cohort schedule: Cohort 1 — 100 fellows starting October 19, 2026; Cohort 2 — ~450 fellows in January 2027; Cohort 3 — ~450 fellows in August 2027.
• Applications close: July 17, 2026 for Cohort 1.
• Host support: Each of the ~400 participating nonprofits receives a $10,000 grant and free Claude credits for the duration of the placement.

What fellows actually build

Fellows are not generic "AI helpers" — each placement has a defined scope. Published examples from early host partners include: building a multilingual case-intake Claude workflow for a legal-aid organisation; creating a grant-writing assistant for a community-health network; and automating donor-reporting pipelines for a conservation nonprofit. CodePath's training curriculum covers prompt engineering, CLAUDE.md configuration, MCP connector setup, and basic agentic workflow design — exactly the skills a non-technical nonprofit staffer needs to become a competent Claude deployer.

🧭 Code with Claude Tokyo Workshop Materials Now Open-Source at anthropics/cwc-workshops

As promised during the Tokyo closing keynote, Anthropic has published all Code with Claude 2026 workshop materials in the anthropics/cwc-workshops GitHub repository under an Apache 2.0 licence. The repo covers every hands-on session from the San Francisco, London, and Tokyo stops of the conference — five complete workshop packages, each with a problem statement, starter code, worked solution, and a skill or MCP server you can drop straight into your own Claude Code setup.

What's in the repository

• rightmodel/ — Picking the Right Model: a Claude Code SKILL that audits an existing LLM eval suite, sweeps it across models and inference parameters, and outputs a cost/quality trade-off table. Immediately useful for teams deciding when to use Fable 5 vs Opus 4.8 vs Sonnet 4.6.
• agent-decomposition/ — Compose Multi-Agent Systems with Skills and MCP: deconstructs a 400-line monolithic agent prompt into modular skills + a code-execution tool + callable sub-agents on Claude Managed Agents. Demonstrates the "why" behind decomposition with before/after token and error-rate data.
• how-we-claude-code/ — How We Claude Code: the written companion to Jason Schwartz's Tokyo Day 2 session. Three-phase walkthrough — CLAUDE.md setup, custom command library bootstrapping, and session-level hook configuration — with annotated examples from Anthropic's internal engineering practice.
• ship-your-first-managed-agent/ — Ship Your First Managed Agent: an end-to-end Streamlit incident-dashboard with an offline SRE Agent chat panel. Good reference for teams adding agent chat to an existing web app without rebuilding the UI from scratch.
• agent-battle/ — Agent Battle: the 45-minute competition format used at all three conference stops. A self-contained benchmark scenario you can use to evaluate competing agent configurations on a fixed task.

The repository also includes the slide decks (PDF) and speaker notes for each session. All workshop code targets Claude Code v2.1.170+ and the Managed Agents API as of June 2026; both the CLAUDE.md in each sub-directory and the top-level README flag any version dependencies.

🧭 Claude Code Gets Safe Mode and Fallback Model Chains — Two Features Every Production Team Needs

Claude Code versions 2.1.166–2.1.169 (shipped June 6–8) introduced two features that quietly solve the most common pain points teams hit when running Claude Code unattended: a --safe-mode flag that isolates configuration problems during debugging, and fallback model chains that keep a production pipeline alive through a transient model overload. Neither feature was headlined at the Tokyo conference, but both are worth understanding before you deploy agents into CI or background worker contexts.

Safe mode — turning "is it my config?" into a one-flag answer

Starting a Claude Code session with claude --safe-mode (or setting CLAUDE_CODE_SAFE_MODE=1) launches a completely clean session: no CLAUDE.md files are loaded, no plugins or skills activate, no hooks fire, and no MCP servers connect. The environment variable alternative is useful for injecting safe mode into background workers without changing your invocation script.

# One-off troubleshooting session
claude --safe-mode

# Or permanently in a test environment
export CLAUDE_CODE_SAFE_MODE=1
claude

# Partial safe mode — disable only hooks and MCP, keep CLAUDE.md
claude --safe-mode=hooks,mcp

The last form (comma-separated component list) is new in 2.1.169. Supported tokens are claude-md, plugins, hooks, mcp, and skills — useful for cases where you want to rule out a specific subsystem without going fully clean-slate.

Fallback model chains — surviving a transient overload

Claude Code 2.1.167 introduced a fallbackModel setting (in settings.json or per-session via --fallback-model) that specifies up to three backup models. If the primary model returns an overload error, Claude Code silently retries the current turn on the first backup. The session reverts to the primary model on the next turn — the fallback is scoped to a single turn, not the entire session.

// settings.json
{
  "model": "claude-fable-5-20261009",
  "fallbackModel": [
    "claude-opus-4-8-20260530",
    "claude-sonnet-4-6-20251022"
  ]
}

Also in this release window

• /cd {path} — moves the active session to a new working directory without breaking the prompt cache. Previously, changing directory mid-session required a full restart and re-warmup.
• Post-session hook (self-hosted runner only) — runs after the session ends and before the workspace is deleted; useful for snapshotting uncommitted work or exporting session logs to a central store.
• MCP policy enforcement tightened — Claude Code now rejects MCP tool calls that request filesystem access outside the declared workspace root, even if the MCP server's declared scope permits it.