BlogParallel agents
Do I need an account or cloud to run coding agents in parallel?
No. Running several agents in parallel only needs terminals and git worktrees on your own machine. You do need an account with each agent's provider, and the agents send your prompts and code context to that provider, because that's where the model runs. Whether a manager app adds its own account or servers on top depends on the app.
So the useful question is not "local or cloud" but "who receives what". Here's how to answer it, as of October 2026.
What leaves your machine no matter which tool you use
The agent itself talks to its vendor.
Claude Code runs locally. To use the model, it sends data over the network, including all user prompts and model outputs, encrypted with TLS 1.2 or later. It also sends operational metrics, which never include code, prompts or file paths. To opt out:
export DISABLE_TELEMETRY=1
# or turn off all non-essential traffic at once
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1
Session transcripts are stored locally under ~/.claude/projects/ for 30 days by default (cleanupPeriodDays).
Codex CLI signs in with ChatGPT or another method on first run. It saves local session transcripts under CODEX_HOME, for example ~/.codex/history.jsonl. To stop that, in ~/.codex/config.toml:
[history]
persistence = "none"
Its default workspace-write sandbox keeps network access off for the commands it runs, unless you enable it.
Gemini CLI collects usage statistics by default. To turn that off, in ~/.gemini/settings.json:
{ "privacy": { "usageStatisticsEnabled": false } }
None of these switches stops the agent from sending your prompts and the files it reads to the model. That's how it works.
The fully local setup
You can run five agents without any account beyond your agent providers:
git worktree add ../app-auth -b auth
git worktree add ../app-search -b search
# one terminal tab per worktree
cd ../app-auth && claude
cd ../app-search && codex
Worktrees, terminals and the hooks or notification settings each CLI offers are all local. See running multiple Claude Code agents in parallel for the full layout.
What the manager apps add
Checked on each tool's own page or repository:
| Tool | Its own account or server | Notes |
|---|---|---|
| Claude Squad | None mentioned | Terminal app on tmux and git worktrees. Needs tmux and gh. |
| cmux | None needed to use it | A macOS terminal, free. Optional paid backing for early features like Cloud VMs. |
| Emdash | No Emdash servers for code or chats | Local-first, state in local SQLite. Telemetry is optional: TELEMETRY_ENABLED=false. |
| Conductor | Local workspaces: model traffic goes straight to the provider, session data on your device | Requires GitHub and an agent login. Analytics are stored on PostHog's servers. Cloud workspaces (paid plans) store session inputs and outputs on Conductor's servers. |
| Superset | Local workspaces on the free plan | Remote access, mobile and cloud workspaces are paid features, and the mobile app signs in with a Superset account. |
The pattern: the local mode of most tools keeps code between you and your agent provider. Cloud workspaces, where they exist, are an opt-in paid feature, and by design they store your sessions on the vendor's infrastructure.
Questions to ask any tool
- Does it need an account to start, or only for some features?
- Does it run agents on its servers, or only on your machine?
- Does it store your prompts, transcripts or code anywhere but your disk?
- What telemetry does it send, and how do you turn it off?
- Which third parties does it talk to: GitHub, analytics, a licence server?
The answers are usually on a privacy or FAQ page. If they're not, that's an answer too.
Limits of "local"
- Local isn't the same as private. Your agent's provider still receives prompts and the code context the agent reads.
- Integrations see what you send. Connecting GitHub means GitHub receives the requests, wherever the tool runs.
- Defaults vary. Gemini CLI's usage statistics are on by default, Claude Code's metrics are on by default with the Anthropic API. Check each one.
- Local sessions stop with your Mac. Conductor's own FAQ notes that local sessions end when the machine shuts off or the app closes. That's the trade-off cloud workspaces exist for.
With Kanbell
Kanbell is a macOS app for Apple Silicon. There is no Kanbell account and no Kanbell server.
- What it talks to: a licence check with Polar, and GitHub if you connect it, for example to use a repository's Issues as the board. Nothing else.
- What still leaves your machine: whatever your agents send to their providers. Kanbell runs
claude,codex,geminiand other CLIs in real terminals, and they behave exactly as they do without it. - Local by design: Codex CLI usage is read locally on your Mac. The board can be fully local. Notes per project live in the repository or stay private.
You get one window per project, agent state for Claude Code, Codex CLI and Gemini CLI, an orange dot and a macOS notification when one waits, and ⌘⇧J to jump to it. Kanbell doesn't run agents in the cloud, so it has no cloud workspaces either.
FAQ
Does Claude Code send my code to Anthropic?
Claude Code runs locally, but to use the model it sends data over the network, including all user prompts and model outputs, encrypted with TLS. Its operational metrics never include code, prompts or file paths, and DISABLE_TELEMETRY=1 turns them off.
Does Conductor send my code to its servers?
Per Conductor's privacy page, in local workspaces model traffic goes straight to your model provider and session data stays on your device. In Cloud workspaces, Conductor stores session inputs and outputs on its servers.
Can I keep Codex from saving history?
Yes. Codex saves local transcripts under CODEX_HOME by default. Set persistence = "none" in the [history] table of config.toml to turn that off.
Is a local-first tool private?
It keeps its own data on your machine. Your agents still send code context to their providers, and integrations like GitHub receive what you send them.