Keep your source code and prompts on your Mac
No row establishes the complete task yet. LM Studio documents offline local chats, document processing and local inference after the model is downloaded, and Free includes local models at $0. It is a runtime rather than a repository editor: the third-party editing tool’s full prompt construction, execution and network behavior still need a supported on-device configuration. Ollama also documents local inference and ways to connect coding agents, but those pages do not establish that every agent action stays on this Mac. Standard Cursor and Claude Code workflows send requests to remote infrastructure; Devin’s complete data path and a qualifying single-device Tabnine configuration remain unknown. Documented local runtime behavior is useful evidence, while no training, privacy mode and zero retention do not by themselves meet the no-transmission rule.
Latest source row read 2026-11-02
The task
Edit an existing private repository on a Mac for up to US$20 a month. Sending source code or prompts to any remote machine, including a company-owned server, is prohibited.
Candidates
Meets the documented conditions is not the same as a completed task: it means the conditions listed in the official pages are satisfied, and the limits below still apply.
LM Studio (local models, Free plan)
Unconfirmed; not recommended yetThis month: US$0 on the Free plan, which the vendor lists as including local model support: within budget.
The vendor documents the runtime side: with model files downloaded, LM Studio can run chats, document retrieval and local inference offline; its server can answer on localhost or the local network. Nothing entered in its local chats and document processing is sent off-device according to that page. Free includes local models at $0. LM Studio is not itself a repository editor: another editing tool must be pointed at the local endpoint, and that tool’s complete prompt construction, execution and network behavior were not established here. This is a documented runtime configuration, not a measured end-to-end private-code workflow.
- Our basis is vendor documentation, not captured traffic: we did not block outbound networking and inspect the requests, so read this as the documented configuration rather than as a measurement of ours.
- Keep the local server bound to localhost: the same page offers serving to the local network, and a request answered on another machine is no longer on this device.
- Intel-based Macs are not supported and 16GB+ RAM is recommended, so check the hardware before making this the route.
- The desktop licence covers personal and internal business purposes, and the cloud models are a separate, US-hosted purchase rather than part of the on-device guarantee.
Ollama local models (Free plan)
Unconfirmed; not recommended yetThis month: US$0 on the Free plan, which the vendor lists as including running models locally: within budget. The paid tiers buy cloud credits rather than the local runtime.
The local half is documented, the editing half is not. The privacy policy states that Ollama runs on your local device and that prompts, responses and other content processed locally are not collected, stored, transmitted or accessible to Ollama, and the local server answers on http://localhost:11434 without an API key - so the model and the execution are on this Mac. What is missing is the surface that edits a repository: Ollama is a model runtime, and the vendor pairs it with third-party agents (its own documentation shows Claude Code, Codex CLI and OpenCode being pointed at the local endpoint) without stating that everything such an agent does stays on the device. The same product also sells cloud models, so the model choice sets the data path.
- Our basis is vendor documentation rather than captured traffic, and the agent half is exactly what those pages do not cover.
- Cloud models are sold on the same pricing page and send prompts to Ollama's infrastructure, so the local guarantee has to be read model by model.
- The terms grant no commercial-use permission, the prices carry no currency code or region, and purchased usage credits expire a year after they are added.
Cursor standard AI workflow
Does not meet this taskThis month: The $20/month Pro amount belongs to eligible existing legacy Cursor subscribers. New customers start on The X Subscription; legacy eligibility and a current currency/market-qualified purchase within US$20 were not established for this brief. The remote-request restriction below independently fails the local-only task.
Does not meet the task. Requests pass through Cursor backend even when you supply your own API key, and privacy mode, no-training and zero-data-retention settings carry their own conditions and abuse-review exceptions. Temporary server-side caching is not on-device processing.
- A data-use statement written on 3 September 2026 is the basis here; we did not capture network traffic.
Devin Desktop / Devin Local
Unconfirmed; not recommended yetThis month: Pro is US$20 a month: within budget.
The Desktop FAQ requires backend connectivity. Devin Local’s harness operates on your machine with local files and tools; its Share action uploads a sanitized transcript, with secrets redacted and system prompts/tool definitions removed, subject to organization, sign-in and client conditions. These pages do not establish that model inference and prompt construction stay on this Mac. No traffic capture was performed, so the no-transmission requirement remains unknown.
- Backend connectivity alone does not prove that source code is sent, and we did not capture traffic to settle it.
Tabnine SaaS / private deployment
Unconfirmed; not recommended yetThis month: The 23 September 2026 record described per-user annual subscriptions, but the original pricing URL redirected to a contact page when checked on 2 November; current prices, currency and a US$20 monthly route remain unverified.
SaaS does not meet the task because inference transmits code. A private remote cluster still leaves this Mac, and an air-gapped deployment would need the model endpoint, logs and telemetry, Mac support and a quote checked - a qualifying single-device setup is unconfirmed.
- Self-hosting a server is not the same as running the model on the client device.
Claude Code Pro
Does not meet this taskThis month: Pro is US$20 a month, or US$200 a year: the monthly price is within budget. The US$200 is a separate annual payment, not a monthly option.
Does not meet the task. It runs in the terminal, an IDE or a browser session, but local execution is network inference: prompts and the relevant context go to the model, so source code and prompts do not stay on this Mac. The self-hosted execution environment offered to Team and Enterprise moves command execution onto your own infrastructure without hosting the model, so it does not close that gap either.
- The account type changes training and retention, which is a different question from whether anything is transmitted: consumer Pro follows the model-improvement switch, while commercial accounts are not trained on by default.
- Zero data retention is enabled separately by eligible Enterprise organisations and covers a limited set of features, so it is not a route to on-device processing.
The hard boundary
Local file editing is not local inference. No training is not zero retention, and zero retention is not zero transmission. Turning off indexing, telemetry or sharing one feature does not establish that every AI request runs offline, and a policy requirement cannot be satisfied by a setting whose scope is narrower than the policy.
Next check
First establish a supported configuration, including the repository editor and its price, that keeps the model, prompt construction, retrieval and execution on this same Mac. Binding the model server to localhost is one condition; it does not describe every action of another tool. Once that architecture is documented, an authorised outbound-blocked workflow and traffic inspection can test its behavior. Neither the missing configuration nor an unperformed test should be presented as a completed recommendation.
Where every claim comes from
Each link was opened and read on the date shown. The note describes what that page actually supports.
- LM Studio - pricing ↗
Free $0 including local models, Bionic+ $20 USD a month and Pro $100 USD a month; no region is named for the prices, and the cloud half is US-hosted with Zero Data Retention.
- LM Studio - offline operation ↗
After model files are downloaded, LM Studio can run chats, document RAG and inference offline. Document processing stays local; the server can serve localhost or the local network. This does not document every external repository-editing agent.
- LM Studio - privacy policy ↗
Local runs are not transmitted; cloud requests are processed transiently, not retained, not used for training or any other purpose, and every party works under Zero Data Retention or equivalent terms.
- LM Studio - terms of service ↗
Version August 23, 2026: the licence covers personal and internal business purposes and prohibits sublicensing, service-bureau and software-as-a-service use. The page has no training clause.
- LM Studio - system requirements ↗
macOS needs Apple Silicon and macOS14 or newer; 16GB+RAM is recommended, 8GB may work with small models/modest context, and Intel Macs are unsupported.
- Ollama - privacy policy ↗
Runs on your local device; prompts, responses and locally processed content are not collected, stored, transmitted or accessible to Ollama; cloud-hosted models are processed transiently and not trained on.
- Ollama - pricing ↗
Free includes running models locally; paid plans include cloud usage credits and local hardware running remains unlimited by the vendor wording. The table prints$without a currency code/market; subscription credit pools are not the local runtime.
- Ollama - terms of service ↗
Terms claim no ownership of user inputs/outputs and no model training on them; the text neither grants nor forbids commercial use. Purchased usage credits expire one year after addition. The displayed price’s missing currency/region is a separate pricing-page observation.
- Ollama - quickstart ↗
Cloud models use an API key while local models run without one, the model is downloaded and chatted with on your computer, and the local server answers on http://localhost:11434.
- Ollama - Claude Code integration ↗
Connects a coding agent to local or cloud models through an Anthropic-compatible API, with the manual setup exporting ANTHROPIC_BASE_URL=http://localhost:11434. The page does not state that the agent's traffic stays on the device.
- Cursor - pricing ↗
Historical 23 September price record: Pro $20/month. It does not establish a current new-customer US$20 purchase; use the current and legacy subscription documents for the eligibility distinction.
- Cursor - data use (3 September 2026) ↗
Updated September3, 2026. Requests go through Cursor backend even with a user API key for final prompt building; file contents are temporarily cached on servers. Privacy Mode has no-training/ZDR conditions and abuse-review exceptions; disabling it permits broader use/storage.
- Devin - desktop FAQ ↗
DevinDesktop requires listed backend hosts, including codeium/windsurf services and Cognition app/API hosts. Backend connectivity alone does not identify all uploaded code/prompt content. Historical registered scope read on 2026-09-23, not rechecked in full on 2026-10-15: Desktop connectivity requirements and what sharing a conversation uploads.
- Devin - pricing ↗
Pro is US$20 per month. This price does not establish that source code and prompts stay on the Mac.
- Devin - local execution ↗
DevinLocal runs its harness on the local machine with localfiles/tools. ExplicitShare uploads a sanitizedtranscript with secrets redacted and system prompts/tool definitions removed, subject to organization permission, Devin sign-in and client support. The page does not promise all inference/prompt construction stays on the Mac. Historical registered scope read on 2026-09-23, not rechecked in full on 2026-10-15: What local execution means here; it does not state that inference and prompts stay on the device.
- Tricentis - contact page ↗
The old Tabnine pricing URL returned a 301 redirect to this contact page on 2 November 2026. The contact page did not provide Tabnine plan prices; the redirect does not establish current subscription terms.
- Tabnine - pricing ↗
Per-user amounts without a currency we could verify, presented as annual subscriptions.
- Tabnine - privacy ↗
The client sends some local project code/context to the Tabnine cluster for inference; local-retention claims do not mean inference stays on the Mac. The page separately describes logs/metrics in self-hosted and air-gapped deployments.
- Tabnine - deployment options ↗
SaaS is hosted; private installations are Kubernetes clusters in a customer VPC or customer servers, with air-gapped Enterprise available. This does not establish a supported single-Mac model endpoint/telemetry setup under$20.
- Claude - pricing ↗
Pro at US$20 a month billed monthly, with the US$200 annual figure separate.
- Claude Code - overview ↗
The terminal-first surfaces: VS Code and JetBrains extensions, desktop app and browser cloud sessions.
- Claude Code - data usage ↗
Consumer Free/Pro/Max training follows the improvement setting. Commercial Team/Enterprise/API and other listed commercial accounts are not trained on under commercial terms unless they choose to provide data, such as through the partner program; retention varies by account.
- Claude - zero data retention ↗
Enabled separately by eligible Enterprise organisations, with limits on the features it covers.
- Ollama — Codex CLI integration ↗
The Codex profile can target localhost:11434/v1 and choose local or cloud models. Its web-search requests are executed through Ollama for either model type and require Ollama sign-in; a local model alone does not establish all agent traffic staying on-device.
- Ollama — OpenCode integration ↗
Open Code is a terminal coding agent; Ollama configures local or cloud models and shows a localhost:11434/v1 baseURL. Other existing Open Code settings still apply; the page does not promise complete device-only agent behavior.
- Claude Code — self-hosted execution vs inference ↗
Team/Enterprise self-hosted environments are beta and off by default. They move session execution to customer compute; conversation/prompt/tool-result content still goes to Anthropic, and inference normally uses Anthropic or configured cloud providers. Self-hosted execution does not host the model on this Mac.
- Cursor - models and current subscription ↗
Cursor is sold as part of The X Subscription, covering Grok, Cursor and X. Its included Cursor-model usage pool is shared across Grok and Cursor and refreshes weekly. Third-party models are excluded; Premium and higher use them from prepaid credits at API list prices, with an additional Cursor Token Rate on Teams and Enterprise. The Grok API is billed separately and does not draw from the pool. This differs from legacy Cursor’s two monthly pools and establishes no fixed coding-task count. The current plan table lists Plus at $8/month or $80/year, Premium at $30/month or $300/year, Super at $100/month or $1,000/year, and Ultra at $200/month or $2,000/year, excluding tax. The full regional and account-specific purchase conditions were not established. The Plus feature list excludes Cloud Agents; Premium and higher add them. A generic API eligibility statement must not be used as a promise of Cloud Agents entitlement on every current plan.
- Cursor - legacy plans and migration ↗
New Cursor customers start on The X Subscription. Existing legacy Cursor plans keep their price and features until changed or cancelled, and switching to The X Subscription is permanent. Legacy Pro is listed at $20 per month and old Teams at $40 per user per month; legacy plans use two monthly pools with extra usage billed in arrears, whereas the new subscription uses a weekly pool and prepaid extra usage. Legacy amounts are not presented here as a generally purchasable new-customer offer.