🧭 Claude Code 2.1.224: Self-Hosted Runner, Cross-Session Messaging, and No More Subagent Cap

Released at 04:00 UTC on August 7, Claude Code v2.1.224 ships three features that meaningfully change how enterprises and multi-agent workflows operate: a self-hosted runner command, cross-session messaging between independent Claude Code sessions, and the removal of the 200-subagent spawn cap.

claude self-hosted-runner

Team and Enterprise customers can now run Claude Code web, mobile, and desktop sessions entirely on their own machines or containers. Sessions run inside the customer's network and can reach internal services — CI systems, private registries, on-premise APIs — without routing data through Anthropic's cloud infrastructure. To start a self-hosted runner:

# Start a self-hosted session runner (Team/Enterprise only)
claude self-hosted-runner

# Sessions will route through this machine rather than Anthropic cloud
# Internal services are reachable as normal

Cross-session messaging: SendMessage and ListAgents

Two new tools — SendMessage and ListAgents — allow separate Claude Code sessions to communicate with each other across machines (macOS and Linux). This enables coordinator-worker patterns where a lead session delegates subtasks to specialist sessions and collects their results.

# In the coordinator session — list active sessions this runner knows about
ListAgents()

# Send a message to a named worker session
SendMessage(
  to: "session-id-or-name",
  content: "Analyse the auth module and return a risk summary"
)

Incoming messages are subject to crossSessionInbound policy settings and optional dialogExpiry — so a worker session can decline unrecognised senders or auto-discard stale messages. Both tools use the same permission model as all other Claude Code tools, so org-level policies apply.

200-subagent cap removed

The previous hard limit of 200 subagent spawns per session is gone. Concurrency and depth limits still apply (the session doesn't get infinite parallelism), but long-running pipeline sessions that previously stalled after spawning 200 workers will now continue without interruption. This change pairs naturally with SendMessage: a coordinator session can now orchestrate far larger agent networks than before.

Other 2.1.224 changes worth knowing

• Archive plugin source: Install plugins from HTTPS zip files with optional SHA-256 pinning — useful for air-gapped or internally hosted plugin registries.
• Sandbox credential-masking options: JWT-aware masking and AWS SigV4 re-signing for fine-grained control over which credentials appear in tool outputs.
• Long project path fix: Paths longer than 200 characters were resolving to wrong session directories — now fixed.
• Sandbox trailing-slash bypass fixed: A sandbox filesystem deny rule with a trailing slash could be bypassed — patched.
• Paste data fix: Paste content was occasionally attaching wrong data or losing text entirely — fixed.

🧭 Anthropic Cuts Fable 5 Biology False Positives by 85% — Virology and Toxicology Still Blocked

Anthropic published a detailed post today explaining how it rewrote Claude Fable 5's biology safety classifier after release. The original configuration was triggering excessive fallbacks on routine health and educational queries — routing them to a less capable model rather than answering them directly. The update cuts those fallbacks by approximately 85% across all product surfaces.

What changed

The team rewrote the classifier's ruleset to better distinguish genuinely harmful requests from benign ones, collected expert feedback on edge cases, and retrained the system on updated data. The classifier now permits:

• Everyday health questions and patient-facing support
• Educational biology content (textbook-level, coursework, general science)
• Clinical support for healthcare professionals

The surface-level breakdown of fallback reductions (compared to the original Fable 5 classifier):

• Claude.ai: ~67% fewer fallbacks
• Cowork: ~55% fewer fallbacks
• Claude Code: ~17% fewer fallbacks
• Claude Platform (API): ~7% fewer fallbacks

What remains restricted

Fable 5 continues to decline requests in dual-use biological domains where the model could meaningfully uplift a malicious actor:

• Virology: pathogen enhancement, gain-of-function techniques
• Toxicology: synthesis of toxic agents at weaponisation scale
• Molecular design: de novo design of biological agents with harmful potential

Anthropic also mentions it is building trusted-access pathways for legitimate frontier researchers who need to work in the still-restricted domains. No timeline or mechanism was given, but the framing suggests a credentialled-researcher programme similar to the External Researcher Access Program for deprecated models.

🧭 Mid-Conversation Tool Changes: Add or Remove Tools Between Turns Without Breaking Cache

Released as part of the Claude Opus 5 launch on July 24 and now available on four models, mid-conversation tool changes let you dynamically add or remove tools between turns of a live session while preserving prompt-cache hits. The beta header is mid-conversation-tool-changes-2026-07-01.

Supported models

• Claude Fable 5 (claude-fable-5)
• Claude Mythos 5 (claude-mythos-5)
• Claude Opus 4.8 (claude-opus-4-8)
• Claude Opus 5 (claude-opus-5)

How it works

Instead of re-sending the full tools array on every turn (which breaks the prompt cache), you inject tool_addition and tool_removal content blocks in a role: "system" message. The model receives an incremental diff of its available tools rather than a complete replacement.

# Python example — add a new tool after the first turn
response = client.messages.create(
    model="claude-fable-5",
    betas=["mid-conversation-tool-changes-2026-07-01"],
    messages=[
        {"role": "user", "content": "Analyse this dataset."},
        {"role": "assistant", "content": assistant_turn_1},
        {
            "role": "system",
            "content": [
                {
                    "type": "tool_addition",
                    "tool": {
                        "name": "run_sql",
                        "description": "Execute a SQL query against the analytics DB.",
                        "input_schema": { ... }
                    }
                }
            ]
        },
        {"role": "user", "content": "Now run a query to find outliers."}
    ]
)