Ivy Tendril
Run, supervise, and review coding agents in isolated Git worktrees.
Per-dimension scores and reasoning
The README places agents in isolated Git worktrees and describes review, approval, and merge before changes reach the main branch, providing substantive evidence for constrained code impact, user confirmation, and limited recovery. It also discloses major external surfaces such as tunnels, voice/file input, and GitHub or webhook integrations. Deductions apply because there is no permission inventory, network or telemetry data-flow account, credential storage/redaction/retention policy, or complete rollback procedure; starting jobs can directly affect code and workspaces. Lockfile-based installation and fixed tool versions help, but Actions use mutable major-version tags, the Ivy-Framework checkout is not revision-pinned, and the recommended installers pipe downloaded scripts directly into interpreters. Copyright and license text attribute the source to Ivy Interactive, although publisher identity remains independently unverified as stipulated.
The README's plan, execute, retry, and status workflow is broadly consistent with the VS Code bridge and chat tests. Tests cover formatting, compilation, unit suites, message validation, missing-argument guidance, cancellation, and unsupported-operation errors. Deductions apply because these are static workflow and test definitions rather than observed results, and backend CI explicitly excludes end-to-end, Claude integration, retry, and health-check tests. Builds also rely on an unpinned Ivy-Framework checkout and several external toolchains without an offline path or comprehensive dependency-failure recovery guidance.
The material identifies developers using CLI coding agents and covers desktop, headless-server, remote/mobile monitoring, review, GitHub issue, and multiple editor/agent scenarios. Slash commands, job types, and bridge messages have concrete parameter validation with thorough trigger-focused tests. Deductions apply because broad claims such as support for any CLI agent and all three desktop platforms are not comprehensively substantiated here, and unsupported agent behavior, OS-specific differences, resource requirements, and environmental boundaries are not clearly catalogued.
The README clearly organizes features, supported agents, skill installation, launch modes, and support channels, with commands and examples for several environments. CreatePlan, ExecutePlan, RetryPlan, and UpdatePlan naming and argument forms are reasonably consistent in tests. The complete FSL-1.1-ALv2 text clearly states present restrictions and the future Apache-2.0 grant, fully supporting the license score. Deductions apply for the absence of an FAQ, troubleshooting section, known-limitations list, upgrade/uninstall guidance, and repository changelog. Version information is largely delegated to a release badge, while responsibility is only indicated through copyright, issues, and Discord, without named maintainers, response commitments, or a security-reporting route.
The interface descriptions and tests show usable plan creation, execution, retry, status tables, live events, diffs, worktree navigation, and review output, strongly supporting output usability. Centralizing multi-agent runs, isolated worktrees, and review also offers plausible marginal value over direct terminal use. Deductions apply because the '99% of code,' '10x,' and IDE-replacement benefits are unquantified marketing claims, with no benchmarks or comparison of pricing, latency, resource consumption, operational burden, or security exposure against the claimed benefits.
Several concrete capabilities are traceable to tests, including message validation, exact command construction, missing-input feedback, event streaming, cancellation, and CI steps that correspond to the README workflow, providing moderate cross-file corroboration. Deductions apply because test source and workflow definitions are not execution records; secure tunneling, automated verification gates, full platform coverage, and universal CLI-agent compatibility lack corresponding evidence in the supplied files, and broad marketing assertions are not clearly separated from verified facts.
- The installation instructions pipe remote shell or PowerShell scripts directly into interpreters; download, version-pin, and inspect those scripts before adoption.
- Cloudflare Quick Tunnels, webhooks, voice input, and file attachments expand the network and data-exposure surface, while the supplied material does not explain authentication, encryption boundaries, telemetry, retention, or sensitive-data handling.
- Do not treat test source or CI definitions as evidence that this revision passed execution; several integration and end-to-end categories are explicitly excluded from the backend workflow.
- FSL-1.1-ALv2 presently restricts competing uses and is not immediately Apache-2.0; verify commercial redistribution or similar-service use and the future-license date for each released version.
What does this agent do, and when should you use it?
Ivy Tendril is a desktop application for agent-driven software development, with a headless web-server mode available through `tendril --web`. It runs terminal-based coding agents in isolated Git worktrees so their changes remain separate from the main branch until a developer reviews, approves, and merges them. Its interface combines inline plan annotations, Whisper voice input, file attachments, diff inspection, and automated verification gates. An automated inbox can ingest GitHub Issues and jam.dev bug reports through webhooks and turn Markdown plans into active jobs. Tendril supports macOS, Windows, and Linux and explicitly works with Claude Code, Codex, GitHub Copilot CLI, Gemini CLI, OpenCode, and other agents that run in a terminal.
Tendril launches terminal-based coding agents and places parallel jobs in isolated Git worktrees, keeping the main branch untouched until the resulting changes are reviewed, approved, and merged. Users can type or dictate prompts with built-in Whisper input, attach text files, logs, or documents, and annotate draft plans inline to revise agent goals. Its review flow displays agent-produced diffs and applies automated verification gates before approval. The automated inbox accepts GitHub Issues or jam.dev bug reports through webhooks and turns Markdown plans into active jobs. Cloudflare Quick Tunnels can expose the server for remote monitoring and steering; tendril starts the desktop application, while tendril --web starts the web server without the desktop UI.
- A developer delegating several coding tasks who wants each agent run isolated in its own Git worktree.
- A reviewer who needs to inspect diffs and verification results before agent-written code reaches the main branch.
- An engineer monitoring and redirecting agents on a remote server from a phone or another computer.
- A maintenance team converting incoming GitHub Issues or jam.dev bug reports into executable jobs through webhooks.
- A user who prefers dictating requirements or attaching logs and documents to a coding prompt.
- A team using Claude Code, Codex, Copilot CLI, Gemini CLI, OpenCode, or another terminal agent that wants one supervisory workflow.
What are this agent's strengths and limitations?
- Parallel jobs run in isolated Git worktrees, providing a concrete boundary between agent changes and the main branch.
- The workflow is not tied to one model provider and is documented to accept any agent that runs in a terminal.
- Plan annotation, diff review, and automated verification gates are integrated into the same supervisory workflow.
- It supports desktop operation, a headless web mode, and remote supervision through Cloudflare Quick Tunnels.
- GitHub Issues and jam.dev reports can enter the workflow automatically as active jobs through webhooks.
- The core workflow depends on Git, local filesystem access, and terminal-based agents, so it is not designed around browser-only chat products.
- Remote tunneling and webhook intake require network access and introduce dependencies on services such as Cloudflare, GitHub, or jam.dev.
- The supplied material does not explain the security model, resource isolation, failure recovery, or the implementation of verification gates.
- The README describes an FSL-1.1-ALv2 source-available license, while the repository metadata reports NOASSERTION; adopters should inspect the actual LICENSE terms.
- Pricing, team authorization, enterprise deployment, and large-scale concurrency limits are not documented in the supplied material.
How do you install or deploy this agent?
Download a standalone macOS .pkg, Linux .AppImage, or Windows .exe from GitHub Releases. On macOS or Linux, the documented quick install is curl -sSf https://cdn.ivy.app/install-tendril.sh | sh. On Windows PowerShell, run irm https://cdn.ivy.app/install-tendril.ps1 | iex. To add the official agent skills, run npx skills add ivy-interactive/ivy-tendril; a single skill can be installed with npx skills add ivy-interactive/ivy-tendril --skill tendril-debug-plan. No credential is documented for the basic launch, and the supplied material does not specify account or authorization setup for GitHub webhooks, jam.dev, or remote tunnels.
How do you use this agent?
Run tendril to open the desktop application, or tendril --web to start the headless web-server mode. Select a coding agent that runs in a terminal and execute work in isolated worktrees; use Tendril to monitor runs, annotate plans, inspect diffs and verification gates, and decide whether to approve and merge the changes. Use the Cloudflare Quick Tunnels feature when remote access is needed, or configure GitHub Issues and jam.dev webhooks for automated job intake. For Codex skills, run codex plugin marketplace add ivy-interactive/ivy-tendril followed by codex plugin add tendril-skills@tendril-skills. For Claude Code, run /plugin marketplace add ivy-interactive/ivy-tendril and then /plugin install tendril-skills@ivy-tendril.
How does this agent compare with similar options?
Tendril is presented as a replacement for the traditional IDE in an agent-driven workflow, but it is not a replacement for Claude Code, Codex, Copilot CLI, Gemini CLI, or OpenCode as coding engines. It sits above those terminal agents and adds worktree isolation, plan revision, run supervision, diff review, verification, and merge decisions, whereas a traditional IDE is primarily organized around direct human code editing.
FAQ
Is Tendril locked to one model or agent provider?
Will it automatically merge agent changes into the main branch?
Can it run on a server without a desktop interface?
tendril --web starts a web server without the desktop UI, and Cloudflare Quick Tunnels can be used for remote monitoring and steering.