Skip to content
Independent runtime research · snapshot 2.1.177

A field atlas of one compiled agent runtime.

Follow Claude Code from native container to provider stream, tool execution, persistence, extensions, and trust boundaries—then continue from each claim into committed evidence and independently authored reconstruction contracts.

Open the system map Choose by goal
Artifact address strip 0x00000000 → 0x0D6B20A0
Prefix · 72,368,128 B Bun section @ 0x04504000 · 150,764,738 B Suffix · 1,991,646 B Open derived topology

Independent research; not affiliated with or endorsed by Anthropic. The atlas publishes evidence summaries and explanatory contracts, not the executable or recovered proprietary source. Read the publication boundary.

Trace Follow a tool call Startup → model stream → permission decision → local execution → tool result. Extend Choose an extension surface Compare instructions, skills, agents, hooks, plugins, MCP, and headless integration. Secure Audit a trust boundary See what crosses the workspace, provider, extension, persistence, and update boundaries. Verify Follow a claim to proof Move from prose to claim ID, sanitized evidence, binary anchor, and reconstructed contract.

Evidence posture

Every non-trivial statement carries one of three confidence classes. The visual language is deliberately structural: text and shape identify the class, while color is only a second signal.

Observed 32 artifact or runtime facts Derived 52 bounded interpretations Hypothesis 1 explicit open model
Observed

The active launcher resolves to a signed arm64 Mach-O at ~/.local/share/claude/versions/2.1.177. Its SHA-256 is eb0730351be2f02b482b1855870f5877489085aac86b0c4c1db4e458d9e40ed9. Artifact provenance

Observed

The executable contains one __BUN,__bun section with an 11-module graph: one large JavaScript entry module, five native-binding loaders, and five matching N-API modules. Binary inventory

Observed

Isolated probes exercised provider streaming, a three-request Read → Bash loop, session persistence, settings precedence, concurrent hooks, MCP stdio, extension discovery, permissions, and sandbox containment. Browse runtime observations

Derived

The executable is best modeled as a local orchestration runtime around a remote model service. The client owns context assembly, extension discovery, permissions, tool execution, persistence, integrations, and transport selection.

System at a glance

flowchart LR
    accTitle: Claude Code internals - System at a glance
    accDescr: Diagram showing system at a glance in the Claude Code internals section.
    User["User or SDK client"] --> CLI["CLI and protocol adapters"]
    CLI --> Config["Configuration and extension discovery"]
    Config --> Loop["Agent turn engine"]
    Loop --> Provider["Provider transport"]
    Provider --> Model["Remote model service"]
    Model --> Loop
    Loop --> Policy["Permission and sandbox policy"]
    Policy --> Tools["Local tools and integrations"]
    Tools --> FS["Workspace, shell, network, MCP, IDE"]
    Loop --> Store["Transcripts, tasks, memory, checkpoints"]
    Hooks["Hooks, skills, agents, plugins"] --> Config
    Hooks --> Loop

Continue with the one-page boundary map, choose an audience-specific route, or inspect the evidence-to-code index.

Deliberate omissions

The project does not publish the native executable, extracted JavaScript, native add-ons, deminified function bodies, authentication material, local configuration, transcripts, or private debug data. Reconstructed modules are independently authored explanatory contracts; they may be incomplete and are not represented as Anthropic's original source tree.