Verde
A native tiling workspace that runs Codex, Claude Code, Cursor and other coding agents side by side, local-first and MIT-licensed.
- Source repo
- JonathanRiche/verde
- Stars
- ★ 36
- Last updated
- 1d ago
- License
- MIT
- Primary language
- Zig
- FA score
- 46/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- Developers running a Claude Code task and a Codex task against the same repo who want both conversations visible side by side.
- Not a fit if
- Teams that want a browser-only tool with no local CLI installs
- Single-agent users who don't need a multi-pane workspace
- Website
- verdeai.dev ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-20latest release v0.1.129 (2026-11-10)
- Source review
- 46/100 · Major gaps
What does this agent do, and when should you use it?
Verde is a native desktop workspace written in Zig 0.16 that tiles GUI chats, terminal TUIs, a file tree and an embedded browser for several coding agents in one window. It ships no model of its own: it drives the provider CLIs already installed and logged in on your machine — Codex, Claude Code, OpenCode, Cursor, Pi, FX, Grok Build and Amp (TUI only) — so it reuses existing subscriptions without new API keys or a hosted relay. Execution is owned by a headless session daemon: chat turns and terminal PTYs live there, while the GUI, CLI, MCP bridge, web client and mobile apps attach as clients, so agent work survives closing the window or upgrading. It is built on the in-tree Palette Zig GUI framework with Ghostty/libghostty-vt as terminal engine, WPE WebKit/WKWebView/WebView2 as the native webview and SQLite for local state. A verde live local IPC surface and a verde mcp endpoint make it scriptable from shells and other agents.
Verde reads a workspace-root verde.toml to declare managed processes and agent launch profiles (for example [processes.web] with command = "npm run dev", or [agents.codex] with provider, revive and hooks) and hands those processes to the session daemon. It invokes locally installed CLIs: Codex through codex app-server, Claude Code through the Claude Agent SDK, OpenCode through opencode serve, Cursor/FX/Grok over ACP (agent acp, fx acp, grok agent stdio), Pi through pi --mode rpc, and Amp inside an embedded Ghostty pane. Users send prompts from GUI chats or terminal TUIs, can approve a command, redirect a running turn or queue a follow-up mid-turn, and the built-in browser's Design Mode lets them click an element or draw a region so the visual context reaches the agent. A Changes panel lists changed files per chat, while verde live and verde mcp let scripts and other agents open panes, send prompts, read transcripts, write to terminals and open the browser. Transcripts, drafts and image attachments are stored in local state.sqlite.
- Developers running a Claude Code task and a Codex task against the same repo who want both conversations visible side by side.
- Engineers who prefer agent TUIs in a real terminal but want a tiling layout, persistent zoom and a project file tree around them.
- Frontend developers who want the agent to see the local dev server page and point at a specific element to change.
- Users who want to approve commands, steer a running turn and push a diff from a phone after leaving the desk.
- Teams scripting panes and prompts from shell hooks or automation via
verde liveandverde mcp. - Omarchy users who want the workspace to follow the active
colors.tomltheme automatically.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
curl -fsSL https://verdeai.dev/install.sh | shirm https://verdeai.dev/install.ps1 | iex
- Config files
verde.toml- Install notes cover
- Windows · macOS · Linux
How do you install or deploy this agent?
On Linux/macOS use the official install script:
curl -fsSL https://verdeai.dev/install.sh | shOn Windows x64, from PowerShell:
irm https://verdeai.dev/install.ps1 | iexOr take a package from GitHub Releases: Linux x86_64 tarball extracted and ./install-local.sh run; macOS .dmg/.zip with Verde.app moved into Applications; Windows zip verified via .sha256 then the packaged install.ps1 (per-user, needs the WebView2 Evergreen Runtime). On Arch Linux:
yay -S verde-binBuilding from source needs Zig 0.16.0 via mise from the repo root (bare zig build is explicitly discouraged):
mise install
mise run setup
mise run dev-build
mise run devVerde has no bundled model, so install and sign in to at least one provider CLI first, e.g. codex login, agent login for Cursor, fx login, grok login, with Node on PATH for Claude Code.
How do you use this agent?
Start with verde in a terminal, then:
- Import a project via the sidebar
+or by right-clicking the rail; threads and agents run against that workspace root. - Click the workspace pencil, or enable prefix mode and press
Ctrl+Bthenc, to create a chat; pick provider, model and run permissions, then send a prompt. - Press
Ctrl+Shift+Bto tile the native webview and preview your dev server. - In prefix mode,
Ctrl+Bthentopens a terminal tab andShift+Ta standalone terminal pane, where Codex/Claude/Grok TUIs can run. Ctrl+Shift+Popens the command palette for threads, panes, workspaces and app commands.
Scripted control from a shell:
verde live status --json
verde live chat open --workspace $WS --provider codex --model gpt-6.1-sol --reasoning high
verde live chat send --pane $PANE --prompt "run the tests and fix failures"
verde live chat followup --pane $PANE --prompt "also update the changelog"
verde live terminal write --focused --text $'cargo test\r'
verde open http://localhost:3000Declaring managed processes and agents (verde.toml):
[processes.web]
command = "npm run dev"
resources = ["port:3000"]
[agents.codex]
provider = "codex"
command = "codex"
revive = "attach_or_create"
hooks = trueImporting a theme:
verde theme import https://verdeai.dev/themes/tokyo-night.jsonWhat are this agent's strengths and limitations?
- Native Zig build on an in-house Palette GUI framework rather than Electron, with low idle cost and smooth long-transcript streaming.
- A headless session daemon owns turns and PTYs, so agent work survives window restarts and upgrades finish current work first.
- Drives already-logged-in provider CLIs locally; no hosted models, no prompt proxying, no telemetry sink.
- Built-in browser with Design Mode turns a clicked element or drawn region into visual context for the agent.
verde livelocal IPC and averde mcpendpoint expose panes, chats and turns to shell hooks and other agents.- Tiling plus a Niri-style scrolling strip, persistent layouts, and automatic theme following under Omarchy.
- Heavy dependence on provider CLIs being installed and authenticated on the host; a missing CLI disables that agent.
- The browser pane needs system WPE WebKit on Linux and the WebView2 Evergreen Runtime on Windows, warned about at install time.
- Source builds explicitly warn against bare
zig buildand require the mise task graph, raising the bar for self-builders. - The iOS and Android companion apps are still in beta, so mobile use should not be treated as stable.
- Documented startup failure modes such as
DatabaseSchemaTooOldneed a separate repair script. - No evidence of localized or Chinese UI in the documentation.
How does this agent compare with similar options?
The README positions Verde against three alternatives: tmux plus CLIs (tiling without a native GUI chat, embedded browser or full scriptable control), hosted agent apps (GUI but with a relay and not local-only), and IDE extensions (agents without comparable tiling and local visual context).
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Verde This agent | 46 · Major gaps | Desktop appFree + model costs | ★ 36 | 1d ago | Zig | Codex · Claude Code |
| MindFlock | 59 · Major gaps | Desktop appFree + model costs | ★ 15 | today | Python | Codex · Claude Code |
| Synara | 57 · Major gaps | Desktop appFree + model costs | ★ 2.3k | 1d ago | TypeScript | Codex · Claude Code |
| PandaCode | 67 · Some gaps | — | ★ 1.3k | today | TypeScript | Codex · Claude Code · Claude.ai |
How does FollowAgents rate this agent?
Why each dimension lost points
README asserts local-first operation, no telemetry sink and no hosted relay, and AGENTS.md sets concrete rules on shared-resource leases, not killing others' processes and never logging clipboard contents — genuine positives. However the product drives multiple coding agents that can write files and run commands, and the permission model, approval flow and sandbox boundaries appear only as prose in the sample with no verifiable implementation evidence, so least_privilege and user_confirmation stay at 1. Data-flow transparency is better documented (no model hosting, no prompt proxying) and earns 2, but mobile pairing and the Connect control plane (OIDC, Postgres, endpoint enrollment) involve remote links and credentials shown only as test fragments, so sensitive_data_handling is 1. Dependency security: package.json pulls @anthropic-ai/claude-agent-sdk and yaml, and README lists third-party components with licenses, but there is no lockfile audit or vulnerability-mitigation evidence, so 1. External effects (curl|sh installer, AUR push, Play upload) have CI evidence but no rollback/revocation story, so rollback is 1. Source attribution is well handled via the third-party component table and font license notes, earning 2.
AGENTS.md documents build targets, test aggregators, lockfile synchronization and release-retry rules in detail, showing engineering discipline. But docs/usage-friction-review.md self-reports crashes, duplicate sends, transcript loss and mobile lag, so self_consistency and failure_messages are each 1. Dependency availability is weak: the app depends on many external CLIs (codex, claude, opencode, cursor, pi, fx, grok, amp) plus system WebView/WPE packages, and the README documents setup but no degradation path, so 1.
Audience and scenarios are clear — parallel agents across desktop, web and mobile, including non-coding workspaces — earning 2. Capability boundaries list the supported-agent matrix and GUI/TUI differences but do not systematically state which actions need approval or where agent capability is limited, so 1. Trigger precision: command palette, prefix keys, CLI and MCP surfaces exist, but the sample shows no handling of trigger conflicts or misfires, so 1. Environment fit covers Linux/macOS/Windows, AUR, WebView2/WPE dependencies and mise builds, earning 2.
Information architecture is clear (README sections, docs table, config cheat sheet), install_notes cover multiple platforms and package managers, and examples_and_faq include a first-five-minutes walkthrough and CLI examples — each 2. Naming stability is 1: many CLI subcommands and config keys with no versioned contract in the sample. known_limitations is 1: limitations surface only indirectly in the friction review, not as a consolidated README section. license is 3: full MIT text plus third-party license notes. versioning_changelog is 1: no CHANGELOG in the sample, only release workflows and tag conventions. maintenance_responsibility is 1: publisher identity is unverified and there is no maintainer or security-contact statement in the repository.
Output usability is 2: the tiling workspace, changes panel and CLI/MCP control are genuinely useful to the target user. Marginal value is 2: it differentiates from tmux+CLI and IDE extensions via the embedded browser Design Mode and the session daemon. Cost-benefit is 1: it requires several external CLIs, system WebView dependencies and a Zig toolchain, and the friction review shows high restart and manual-verification overhead.
claim_traceability is 1: performance claims (sub-millisecond prompt acceptance, low idle cost) have no benchmarks or data. cross_source_corroboration is 1: README, AGENTS.md, the friction review and CI workflows partly corroborate each other on build and release process, but product-capability claims lack code-level support. fact_inference_separation is 1: the friction review clearly separates observations from ideas, which helps, but README marketing statements are interleaved with facts.
- The product drives multiple coding agents with file-write and command-execution capability; the sample shows no verifiable permission model, approval flow or sandbox implementation, so review independently before deployment.
- Installation uses curl|sh and PowerShell irm|iex, and the Connect control plane involves OIDC, Postgres and remote endpoint enrollment; credential and remote-connection boundaries need separate assessment.
- docs/usage-friction-review.md self-reports that a shared-daemon crash takes down every session, plus duplicate sends and transcript loss; stability risk needs empirical confirmation.
- Publisher identity is unverified and the repository has no maintainer or security-contact channel, so the update and vulnerability-response path is unclear.
- It depends on several external CLIs and system WebView/WPE packages; missing any one affects availability, and the sample shows no degradation path.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
Release covering all 323 commits since v0.1.128. Providers and agent control - Add Antigravity (agy) across desktop, daemon, headless, web and mobile, including model discovery, installation, updates, readiness and branding. Thinking-tier models display their base names. - Add Claude Fable 5.2 and Muse Ember support. - Add per-session effort, Fast mode and steering for fx. - Add Update all for outdated provider CLIs in Settings → Providers. - Reopen saved chats with CLI chat present and MCP present_chat, preserving identity, transcript and settings. …View on GitHub ↗ -
Release notes excerpt
Highlights - macOS builds are signed and notarized. Release builds are now signed with a Developer ID (RICHE VENTURES, INC.) and notarized by Apple, so Gatekeeper treats Verde like any other downloaded app. Privacy permissions such as Documents access now carry over to the next update instead of being asked for again after every update. The signature changes from ad-hoc to Developer ID with this release, so macOS asks one last time after you install it (6fbf95e7). …View on GitHub ↗ -
Release notes excerpt
Highlights - The Linux browser pane works with WPE WebKit 2.54 again. Arch's wpewebkit 2.54 dropped WebKit's legacy libwpe API. The v0.1.126 browser helper linked that API and failed to start on updated Arch/Omarchy systems, so the browser pane went blank. Release builds now ship only the WPE Platform backend, which was already the default at runtime (58cc9109, ed0b4122). …View on GitHub ↗
FAQ
Does Verde pay for model usage?
Do agents stop when I close the window?
Can I trigger agents from scripts?
verde live for opening chats, sending prompts, mid-turn follow-ups, terminal writes and opening the browser; verde mcp exposes an equivalent MCP surface.Why is bare zig build discouraged?
--release=safe -Dbrowser-backend=native_webview explicitly.Is there a phone client?
Does Verde work on Windows?
irm https://verdeai.dev/install.ps1 | iex.Is Verde open source?
Is Verde safe to use?
Is Verde still maintained?
Who makes Verde?
What is Verde's official website?
Thanks — we use this to improve the page.