Alethe — Multi-Agent Coding Workspace
A local-first desktop workspace to run, organize, and resume multiple coding agents and shells side by side, replacing terminal-tab chaos with persistent project layouts.
Evidence shows proactive disclosure of network defaults (update checks and usage polling on by default), Remote Control off by default with layered opt-ins, read-only PR review, no stored GitHub token, and backed-up atomic MCP config writes. Deductions: implementations cannot be verified in the provided source; Remote Control uses unencrypted HTTP/WebSocket; production CSP is null; default-on update checks are not a least-privilege default.
CI covers three platforms with lint, format checks, vitest, and cargo test; README claims crash-resilient session resume. Deductions: clippy is continue-on-error (~37 pre-existing warnings); only test setup.ts is visible so coverage depth is unknown; error-message quality has no source evidence.
Clear target scenario (multi-agent, multi-repo development), three platforms, EN/pt-BR localization, native installer probing (npm/WinGet/Scoop etc.). Deductions: trigger precision cannot be assessed (no agent prompt/trigger logic in evidence); Linux/macOS self-described as least tested.
Well-organized docs: README, SECURITY.md, PRIVACY.md, CONTRIBUTING, TRADEMARK, SHOWCASE; install notes include source-build deps per platform; full AGPL-3.0 text; known limitations (unsigned builds, false positives, recovery steps) amply disclosed. Deductions: no CHANGELOG file, version only 1.6.0 in package.; single individual maintainer, and SECURITY.md explicitly promises no response deadlines.
Clear product positioning (multi-agent workspace vs. terminal tabs) with evident marginal value in CLI installation, MCP management, and session handoff. Deductions: output usability (handoff redaction is best effort) and cost-benefit rest on absent code evidence; Remote Control and Spotify integrations enlarge attack surface.
Claims point to checkable files (PRIVACY.md, LICENSE, roadmap checkboxes match shipped features, CI config matches badges). Deductions: only README/SECURITY/package./CI were provided; key files like PRIVACY.md are absent, so security/privacy claims cannot be cross-verified; fact/inference are mostly separated but many claims rest on unattached files.
- Remote Control uses unencrypted HTTP/WebSocket on the LAN; enable only on trusted networks and understand the layered input opt-ins even though it is off by default.
- Windows builds are unsigned and macOS builds are not notarized; verify downloads against official Releases and never bypass security warnings for non-official artifacts.
- Production Tauri config has csp: null (self-disclosed in SECURITY.md); do not treat CSP as a command-authorization boundary when assessing risk.
- Update checks and provider usage polling are on by default — not local-only; consult docs/PRIVACY.md and disable if you need offline operation.
- Context handoff redaction is best effort; manually review the packet before starting the target agent.
- Spotify credentials are stored in local profile files; read the privacy guide before exporting or sharing profile data.
- Early release (1.6.0) with no CHANGELOG and no promised security-response deadlines; enterprises should assess adoption risk independently.
What does this agent do, and when should you use it?
Alethe is a cross-platform (Windows, macOS, Linux) desktop application built with Tauri, Rust, React, and xterm.js, positioned as a multi-agent coding workspace. Every agent runs in a real PTY inside a persistent per-project layout; closing a container does not kill the process, and sessions of Claude Code, Codex, and OpenCode resume after crashes or restarts. It also centrally manages what agents share: installing and updating their CLIs, MCP server configurations (read from and written to Claude Code, Codex, OpenCode, and Antigravity configs with atomic commits), skills, and handoffs of conversations between agents (e.g., a locally redacted context packet from Claude Code to Codex). At the project level it ships a Git panel, worktrees, PR review (read-only AI review with human-only squash merging), and a Graphify code graph served to agents as an MCP server. The project is in early public release under AGPL-3.0-or-later; "local-first" describes workspace persistence, not an offline guarantee — update checks and provider usage polling are on by default.
Alethe organizes work as groups → projects → containers → panes → sub-tabs. Each project is a saved working context; opening it creates a container holding one or more panes. Each pane runs a real PTY (pwsh/bash/zsh or an agent CLI), and multiple agents can live as sub-tabs in the same pane, each with its own PTY, working directory, and session. It installs, updates, and uninstalls agent CLIs (claude, codex, copilot, agy, opencode, mimo, freebuff) by probing for Node, npm, WinGet, Scoop, and Chocolatey, and discovers already-installed CLIs across PATH, registry, npm/pnpm/Volta/fnm/nvm/Bun/Cargo, and more. The MCP tab reads MCP servers configured for each agent, lets you copy a server between agents, search the official registry, and ask each agent to verify reachability; every write is backed up, re-parsed, and committed atomically. For context: Claude Code, Codex, and OpenCode sessions resume; "Recent chats" lists conversations by a pane's working directory and reopens them; a Claude Code conversation can be handed to Codex (and back) via a locally redacted context packet. It also provides a per-project Git panel (status, stage, commit, branches, diffs, worktrees), PR discovery via the local gh CLI with read-only AI review, a RAM readout with per-terminal disable and group suspension, content panes (file explorer, Markdown, diffs, images, embedded browser), 14 UI/terminal themes, EN and pt-BR localization, local backup export/import, and a QR-paired LAN Remote Control web view (off by default).
- A developer juggling several repos at once: run Claude Code and Codex in parallel across grouped project containers; close a container and the processes keep running until you return.
- A team fighting MCP config drift: use the MCP tab to see which agents use which servers on a machine and copy a server's config from one agent to another with atomic writes.
- Cross-agent collaboration: hand a Claude Code conversation to Codex through a redacted context packet instead of copy-pasting the thread.
- A PR author: run a read-only AI review on a GitHub PR inside its isolated worktree, with squash-merging remaining an explicit human action.
- A heavy multi-agent user managing memory: watch the title-bar RAM readout and disable terminals or suspend whole groups to reclaim memory.
- Mobile follow-up: pair by QR code on a trusted LAN and answer agent questions from your phone via Remote Control.
What are this agent's strengths and limitations?
- Real PTYs with session resume: closing a container keeps processes alive; Claude Code, Codex, and OpenCode sessions recover after crashes or restarts, and scrollback is persisted per PTY.
- Centralized MCP and skills management: cross-agent config reading, copying, and verification with atomic, backed-up writes — directly addressing config drift across agents.
- Cross-agent handoff: a locally redacted context packet moves a Claude Code conversation to Codex (and back) without manual copy-paste.
- Governance built in: PR AI review is read-only; squash merging requires explicit human action and is blocked on moved head SHA, drafts, or GitHub-reported conflicts; no GitHub token is stored — auth is delegated to gh.
- Early public release: Windows builds are unsigned (flagged by Defender) and macOS builds are not notarized; both are still on the roadmap.
- Not fully offline: update checks and provider usage polling are on by default, and Remote Control uses unencrypted HTTP/WebSocket on the LAN, so it should only run on trusted networks.
- Uneven platform maturity: the README states Linux and macOS are the least-tested platforms.
- Possible future proprietary hosting: first-party cloud sync/backup may be offered separately; today only manual GitHub Gist Sync exists.
How do you install or deploy this agent?
Two options. (1) Installers: download from GitHub Releases (https://github.com/Kc1t/alethe-agents/releases) for Windows/macOS/Linux. Caveats: Windows builds are unsigned — Defender may flag alethe.exe as Trojan:Win32/Bearfoos.A!ml; restore it via Virus & threat protection → Protection history and exclude %LOCALAPPDATA%\Alethe. macOS builds are unsigned/un-notarized — right-click the app and choose Open. (2) From source: git clone https://github.com/Kc1t/alethe-agents.git && cd alethe-agents && npm install && npm run app. Requirements: Node.js 18+, Rust stable; Visual Studio Build Tools on Windows; on Linux run sudo apt install -y libwebkit2gtk-4.1-dev libappindicator3-dev librsvg2-dev patchelf.
How do you use this agent?
Launch the desktop app, create projects and groups, and pick an agent or shell for each pane; missing CLIs can be installed in-app. Optionally install the alethe terminal command via Settings ▸ Integrations ▸ Terminal command: alethe opens the current folder as a project and alethe ~/some/project opens a given folder (existing projects are brought into the workspace, not duplicated). Day-to-day: manage shared MCP servers in the MCP tab and per-agent skills in the Skills tab; use the Git panel for staging, commits, diffs, and worktrees; find and review PRs through the PR panel backed by the local gh CLI. Optional capabilities (Remote Control, Spotify Now Playing, GitHub Gist Sync) are enabled explicitly in settings.
How does this agent compare with similar options?
The README positions Alethe as "closer to Obsidian than to a maximalist IDE." Versus a plain terminal multiplexer, its differentiator is the multi-agent focus: persistent per-project layouts, MCP/skills management, session resume, and cross-agent handoffs. It does require the underlying agent CLIs (claude, codex, etc.) to be installed — by you or through the app — and provides no model inference itself.