Detailed report
Jump to a section
Diri combines existing agent subscriptions, terminals, SSH execution and persistent sessions. It suits a solo engineer using a Mac to control local and remote work. Laptop handoff and team administration are less established.
This assessment uses sources checked on November 13, 2026. Most official documentation is undated. Dated releases and reports are identified below.
Product and workflow
The product is lowercase diri, at diri.sh, maintained in cristicretu/diri.
Diri lets you keep your preferred agents and infrastructure, while replacing the task of watching many terminals with one native workspace. It identifies sessions that need attention and keeps execution separate from the window. The documented architecture comprises the app, a background Engine, and a terminal-owning “Holder” process for each session. Closing the interface therefore does not end the work. Agents can also coordinate helpers through Diri’s built-in MCP interface, which lets agents call external tools. Architecture documentation
Agents, models, and billing
Diri wraps command-line agents. It does not supply a proprietary model. The catalog lists 21 agents plus a shell, while the homepage advertises 22 definitions. These include Claude Code, Codex, Cursor, Gemini, Aider, Amp, Antigravity, Cline, Copilot, Devin, Droid, Grok, Hermes, Kilo Code, Kimi, Kiro, Maki, OpenCode, Pi, Qoder, and WhipCode. Additional terminal tools can use custom JSON manifests.
Integration depth varies substantially. Claude Code has hook-based status. Codex has exact conversation resume and fork support. Cursor and Gemini support exact resume, while several others resume only the latest conversation in a folder or cannot resume. Automatic Diri MCP configuration is documented for Claude Code, Codex, and Cursor. Model and provider availability follows the underlying agent. Agent capability matrix
Diri itself costs $0, with no published seat subscription or bundled inference allowance. You pay separately for your agents, model APIs, and infrastructure. Existing Claude/ChatGPT subscriptions can be used through their agents. API-key configurations depend on the selected agent. Diri is therefore compatible with subscription and bring-your-own-key workflows, rather than providing its own universal billing gateway. Official pricing statement
Claude and Codex account profiles include plan-limit displays and estimated spending. Estimates are not bills, and account switching can interrupt running tools. Accounts documentation
Your servers and disconnected work
This is Diri’s clearest documented feature for your circumstances. Ordinary remote execution uses OpenSSH, including SSH aliases and jump hosts. Diri uploads a verified helper that owns the agent terminal. The agent runs on your server, not a vendor sandbox. Install and authenticate the agent on that server first. Local credentials are not copied.
Supported remote targets are Linux x86-64, Linux ARM64, and Apple-silicon macOS. Intel Macs are excluded as remote targets. Diri tests whether the host preserves detached processes and reports native detach, an existing user supervisor, or a non-persistent warning. Persistent hosts keep working through connection loss and reconnect to the same session. Server reboot stops execution. Only one Diri window controls a session at a time. Remote-host documentation
Local sessions likewise survive app quits and Engine updates, but machine restarts require agent-supported conversation resume. That distinction matters: preserving a running process and reopening its conversation are different guarantees. Session lifecycle
Device handoff and phone access
An optional diri-node service documents moving or forking Claude/Codex conversations between machines, account management, and fleet usage totals. It requires a private Tailscale connection, token enrollment, and systemd user services. Handoff uses checkpoints and provider-native resume. Credentials and selected sensitive files are excluded from transfers. This is more involved than ordinary SSH. It does not establish seamless discovery of every existing session from a fresh second laptop. Treat that specific workflow as unproven until trialed. Node setup and handoff
The iPhone beta can start, monitor, and answer local or preconfigured SSH sessions through a gateway on your Mac over Tailscale. It requires a signed build, rather than an App Store download. The Mac must remain awake with Diri open. Closing its lid disconnects phone access. Pairing grants control over all exposed sessions. I found no documented Android equivalent. iPhone architecture and setup
Workflow and orchestration
Optional Git worktrees give tasks separate checkouts and branches. The review panel supports diffs, staging, commits, and requests for agent review or test suggestions. PR checks, review state, and comments appear beside sessions. Merging a PR opens GitHub. Remote changes are view-only in the panel. Diri’s branch-integration tool supports merge, squash, or cherry-pick, but only between local sessions in the same project. Tests remain commands executed by agents, rather than a documented mandatory quality gate. Worktrees and review
Diri documents agent coordination, with lifecycle issues worth testing. Maintainer issue #107, opened August 13, 2026 and still displayed open, describes ambiguous child completion, unsafe report timing, and lifecycle races. That is historical engineering evidence, not proof every described defect remains in the version available on the research date. It makes completion and interruption behavior essential trial targets. Orchestration issue
Setup, security, and maturity
The primary desktop platform is macOS 15+, Intel or Apple silicon, installed through a notarized DMG or Homebrew. Linux is beta. Documentation mentions Ubuntu x86-64 and ARM64, while the repository’s installation summary still emphasizes x86-64 and warns that Linux packages do not accompany every release. Check actual assets before committing a Linux workstation. Windows remains in progress. Installation
The current license is AGPL-3.0. Older directories and articles still describe Apache-2.0, so those descriptions should not guide redistribution decisions. The repository displayed 924 commits and 491 stars when checked. v0.11.5, released November 9, 2026, includes remote-session resilience and terminal fixes. That supports active development, not established long-term reliability. Repository, release notes
Diri is self-managed but not sandboxed. Agents run with your user privileges. Worktrees do not isolate credentials or network access. Terminal history can contain secrets. Diagnostics upload by default unless disabled. The policy describes identifiers, performance data, and a name defaulting to the macOS login name, with 30-day retention. Self-managed execution still involves diagnostic data leaving the machine. Security and diagnostics
User reports
Independent evidence is sparse. There is no defensible consensus.
An August 5, 2026 source review praised install and updater safeguards and conditionally endorsed personal-Mac installation. It explicitly had not completed a daily-use trial. Its SSH/tmux description also predates current documentation, so its findings need version-specific interpretation. Autopilot Engineer review
Actual user reports identify concrete friction. On October 26, a user reported unexpected private-folder permission prompts caused by home-directory indexing. The thread links an October 29 corrective commit but remains open. On October 14, another requested separate personal/work configuration homes because login profiles still shared provider settings. These are useful cautions about privacy expectations and account separation, not a measure of broader user satisfaction. Issue #826, issue #781
Fit and limits
Choose Diri if you want a Mac-centered workspace for agents on your own infrastructure and accept hands-on configuration. Small teams can adopt it individually. Its own comparison documentation says it lacks team features. Do not assume shared permissions, centralized audit, SSO, or collaborative session ownership. Vendor comparison
The main concern is that good SSH execution does not establish a complete system for control across devices. Avoid making it your sole operational interface before validating laptop handoff, credentials, and failure recovery.
Practical trial
Use a disposable repository on one Mac and one headless Linux server. Run Claude and Codex tasks, disconnect SSH, quit and reopen Diri, and check for duplicate execution or lost input. Then attempt control from your second laptop and phone, test personal/work account separation, and rehearse server reboot recovery.
| Comparison fact | Diri |
|---|---|
| Product cost | $0, agents/models/infrastructure separate |
| Desktop | macOS, Linux beta, Windows pending |
| Headless execution | Own Linux servers over SSH |
| Persistence | Disconnect-aware, reboot requires resume |
| Cross-device | iPhone beta, optional node handoff |
| Isolation | Optional worktrees, no security sandbox |
| Team administration | No documented team feature set |
| Main uncertainty | Reliable laptop-to-laptop session control |
