---
title: "Where Claude's Working State Lives: Local Disk vs. Durable Store"
description: "Compare two places to keep Claude's working state: curated files on your own machine versus a durable session store or isolated cloud computer, and learn which one fits your deployment."
---

# Where Claude's Working State Lives: Local Disk vs. Durable Store

**Type:** [Comparison](https://yomu.fyi/visual-explainers/category/comparison.md)

![A two-column comparison. Left, 'On the host': CLAUDE.md, Skills, Hooks, checkpoints, and local session transcripts. Right, 'Durable or remote': a mirrored session store, an isolated cloud computer, checkpoint rollback, and a network allowlist. A bottom banner gives the decision cue.](https://yomufyi.s3.us-east-1.amazonaws.com/visual-explainers/ce947b6c-3d75-4fd7-9447-9347669458bb/image-f860527e-2390-41cc-9091-677b990d8283.webp)

> Compare two places to keep Claude's working state: curated files on your own machine versus a durable session store or isolated cloud computer, and learn which one fits your deployment.

## Explanation

Claude agents need somewhere to keep both context and session state. One option keeps everything on the host: CLAUDE.md, Skills, Hooks, and checkpoints on local disk, with the Agent SDK writing transcripts to the local filesystem. The other moves that state into a durable store or an isolated cloud computer, so sessions survive disposable instances, host failures, and parallel workers, and can be rolled back or audited. The comparison shows what each choice costs in setup, portability, and control, and gives a decision cue for picking between them.

## FAQ

### When is a local CLAUDE.md and Skills setup enough for a Claude agent?

When the host is long-lived and the work is interactive, files on disk are the simplest option: they travel with the repository and need no extra infrastructure. The Agent SDK also writes session transcripts to the local filesystem by default, so a persistent machine keeps continuity for free. It breaks down when each invocation gets a disposable instance.

### What breaks when Claude's session transcripts live only on local disk?

Stateless deployments create a fresh instance per invocation, so a local transcript is lost and the agent restarts cold. Host failure loses the session, and there is no shared record to audit or review later. Mirroring each turn to an external store, as the MongoDB session adapter does, restores continuity under existing access controls.

### Why give Claude an isolated cloud computer instead of running it on my laptop?

An isolated virtual machine removes the permission prompts that come from running commands on your own machine, and it lets the agent install dependencies and start servers without collisions. Checkpoint snapshots allow rollback, and a DNS-level network allowlist limits where code can send data. The trade-off is provisioning: code and credentials move into the remote environment.

**Tags:** [Agent Memory](https://yomu.fyi/topic/agent-memory.md), [Claude](https://yomu.fyi/topic/claude.md), [Claude Code](https://yomu.fyi/topic/claude-code.md), [Context Engineering](https://yomu.fyi/topic/context-engineering.md), [MCP](https://yomu.fyi/topic/mcp.md)

## Related reading

### [Run Claude Code From Your Phone With Sprites](https://yomu.fyi/post/run-claude-code-from-your-phone-with-sprites.md)
- Company: [Fly.io](https://yomu.fyi/company/fly-io.md)
- Published: Nov 6, 2026
- Type: Tutorial

Sprites are on-demand, hardware-isolated Linux computers that start in under a second, persist code, packages, and state between sessions, and let you checkpoint the entire environment at any time. The sprite-mobile tool provides a chat UI for issuing commands to Claude Code in YOLO mode from a phone over Tailscale, enabling unrestricted and aggressive experimentation. Setup requires one command to create a Sprite and a one-line script that configures Claude CLI, GitHub CLI, Fly.io CLI, Sprites CLI, and Tailscale, sets up HTTPS via Tailscale Serve, and starts Tailnet Gate for waking suspended Sprites. After setup, obtain the public URL with sprite url update –auth public -s cheese\_sandwich and open it in a Tailscale-connected phone’s browser. Sprites automatically sleep after 30 seconds of inactivity, wake in under a second, and incur no charges when idle.


### [Context Engineering Is Just Writing Things Down](https://yomu.fyi/post/context-engineering-is-just-writing-things-down.md)
- Company: [Fly.io](https://yomu.fyi/company/fly-io.md)
- Published: Nov 6, 2026
- Type: Problem & solution

Context engineering is the practice of managing agent memory through curated files on disk, using hybrid search (BM25 and vector search with 70/30 weighting and diversity re-ranking) backed by local SQLite. The post argues work quality depends more on project context than model capability, advocating four mechanisms in Sprite: CLAUDE.md, Skills, Hooks, and Checkpoints. Daily logs are append-only and decay over time, while MEMORY.md only changes when deliberately promoted, representing a stream versus curated memory. The post illustrates context engineering with an Alibi excuse-as-a-service API built on Bun, Hono, and Postgres, where CLAUDE.md defines stack, commands, naming schemes, error format, logging, and test locations. Therefore, an agent can correctly fix a bug by sampling without replacement and write a regression test using bun test because the persistent context provides project knowledge.


### [Closing the Loop: The Missing Ingredient in Agentic Development](https://yomu.fyi/post/closing-the-loop-the-missing-ingredient-in-agentic-development.md)
- Company: [Fly.io](https://yomu.fyi/company/fly-io.md)
- Published: Nov 6, 2026
- Type: Problem & solution

AI code agents that require constant permission prompts create a new form of friction similar to old copy-paste workflows. The Sprite environment solves this by giving agents unrestricted access to a fully isolated computer where they can install dependencies, start servers, and test endpoints without human intervention. Preinstalled tools such as test runners, service logs at known paths, and checkpoint-based rollback enable agents to write code, run it, observe failures, and fix issues autonomously. Additional tools like curl, jq, sqlite3, and network debugging utilities support API testing, database inspection, and network troubleshooting. The result is a tight closed loop where the agent can build, test, and ship without prompting the developer, turning the developer into the idea person rather than middleware.


### [Claude Platform now supports native OpenTelemetry observability in Databricks MLflow](https://yomu.fyi/post/claude-platform-now-supports-native-opentelemetry-observability-in-dat.md)
- Company: [Databricks](https://yomu.fyi/company/databricks.md)
- Author: Yuki Watanabe
- Published: Nov 3, 2026
- Type: Announcement

The Claude Platform now supports native OpenTelemetry export to Databricks MLflow, enabling developers to trace agent executions. Traces record model requests and tool calls as spans, ingested via Zerobus and stored in Unity Catalog tables linked to MLflow experiments. Setting up requires creating an experiment with Unity Catalog trace storage, configuring native export with a table-scoped OAuth token, and running a task to inspect traces. Developers can debug unexpected answers by analyzing traces with Genie Code, which identifies missing context or failed tool calls, and MLflow Issues detects recurring problems across traces. Agent Evaluation compares baseline and revised agents using the same traces to measure improvements or regressions, with domain experts able to review traces via Review Queues.


### [Enterprise-managed auth for MCP: Claude can now issue the identity assertion](https://yomu.fyi/post/enterprise-managed-auth-for-mcp-claude-can-now-issue-the-identity-asse.md)
- Company: [Supabase](https://yomu.fyi/company/supabase.md)
- Author: cemal\_kilic, wenbo
- Published: Nov 3, 2026
- Type: Announcement

Supabase’s MCP server now supports enterprise-managed authorization when an organization’s identity provider lacks the MCP Enterprise-Managed Authorization extension, because Claude can issue the signed identity assertion itself. After a Supabase owner or admin approves Claude and registers it as the issuer, members with matching Supabase SSO identities use Supabase from Claude without a per-member consent screen. Claude checks that the person is a current member of the Claude organization and which email address it uses, but it does not confirm an identity-provider sign-in, check multi-factor authentication, or apply Supabase’s sign-in rules. Supabase verifies the signature against the registered issuer, the member’s matching SSO identity, and the admin authorization before issuing a short-lived access token. Disabling a member in the identity provider no longer cuts off access, so they must be removed from both organizations, and tokens remain valid until expiry.


### [Introducing the MongoDB Claude SDK Session Adapter: Durable Session Storage for Agents](https://yomu.fyi/post/introducing-the-mongodb-claude-sdk-session-adapter-durable-session-sto.md)
- Company: [MongoDB](https://yomu.fyi/company/mongodb.md)
- Published: Oct 26, 2026
- Type: Announcement

The Claude Agent SDK embeds Claude Code’s agent loop into applications, providing tools like web search and MCP connections without manual tool definition. It automatically persists session transcripts to the local file system, but this is unsuitable for stateless deployments like Lambda functions where each invocation is a disposable instance. The MongoDB Claude SDK Session Adapter solves this by implementing the SDK’s SessionStore interface to mirror every conversation turn to MongoDB in real time. This enables cross-environment continuity, resilience to host failures, and auditability under existing access controls, without requiring a separate system for agent memory. The adapter can be configured with a few lines of code, making agent sessions portable across any instance with access to the MongoDB cluster.


### [Give Claude Code a cloud computer with MCP](https://yomu.fyi/post/give-claude-code-a-cloud-computer-with-mcp.md)
- Company: [Boxcompute](https://yomu.fyi/company/boxcompute.md)
- Author: boxy
- Published: Oct 26, 2026
- Type: Tutorial

Claude Code runs commands on your local machine, but tasks like cross-platform builds or heavy compilation benefit from a remote VM. This guide connects Claude Code to a cloud Linux VM through the BoxCompute MCP server, enabling Claude to create, start, and stop VMs, run commands with up to 15-minute background execution, read and write files under /workspace, and generate one-hour public preview links for ports. Setup takes about two minutes: add the MCP server via a CLI command, authenticate through a browser sign-in, then ask Claude in plain language to create a VM and work on it. A small VM costs roughly $0.05 per hour, stops after thirty minutes of idle time, and retains files for later use. The same MCP server works with Codex and Cursor, requiring only an API key file.
