Claude Code and Ivy Tendril are not alternatives. Claude Code is a coding agent from Anthropic that runs in your terminal, reads and edits files, runs commands, and commits. Ivy Tendril is a desktop application that orchestrates coding agents, and Claude Code is one of the agents it runs. When you use both, Claude Code does the work inside each plan, and Tendril supplies what a single terminal session does not: a written plan that a human approves, isolated worktrees for many concurrent runs, a verification step, a review surface for the diff, and cost tracking across everything. If you run one agent at a time and watch it, Claude Code alone is enough. If you want to run ten and review the results, add Tendril.
What each tool is for
Claude Code
Claude Code is Anthropic's command-line agent. You start it in a repository, describe a task, and it works through the codebase: it reads files, proposes edits, runs tests and other commands, and asks permission for actions you have not pre-approved. It reads a CLAUDE.md file for project instructions, supports MCP servers, has hooks that run shell commands at defined points, and can spawn subagents for parallel subtasks. It has a plan mode that lists intended steps before editing. It runs Anthropic's Claude models, available through a Claude subscription or an API key, and through Amazon Bedrock and Google Vertex AI. At the time of writing it is one of the most capable single agents for working in an existing codebase.
Its scope is the session. When you close the terminal, the record of what was planned, what was verified, and what it cost is your shell history and your git log.
Ivy Tendril
Ivy Tendril is a local-first desktop application for macOS, Windows, and Linux, also available headless with tendril --web. It runs a lifecycle the product calls a software factory: a ticket or idea becomes a plan, a human reviews the plan, an agent executes it in a git worktree, tests and lint run, a human reviews the diff, and a pull request is opened. It is free and source-available under the Functional Source License. Pro and Enterprise add team features, on-prem hosting, SSO, verification imports from CI, and support.
Tendril does not include its own model or agent. It drives Claude Code, OpenAI Codex CLI, GitHub Copilot CLI, Google Gemini CLI, OpenCode, or any CLI agent, with your own API keys. The Claude Code integration guide covers setup.
Side by side
| Aspect | Ivy Tendril | Claude Code |
|---|---|---|
| Category | Agent orchestration | Coding agent |
| Runs | On your machine, desktop or headless | In your terminal |
| Models | Any provider you hold a key for | Anthropic Claude models |
| Agents | Claude Code, Codex CLI, Copilot CLI, Gemini CLI, OpenCode, others | Itself |
| Unit of work | Plan, from ticket to pull request | Session |
| Human checkpoints | Plan approval and diff approval | Permission prompts per action, plus plan mode |
| Parallel runs | Many plans, each in its own worktree, one Dashboard | One session per terminal, subagents inside a session |
| Verification | Tests, lint, and diff shown together in Review | Whatever the agent runs during the session |
| Memory | Per-promptware Memory/ that agents update | CLAUDE.md and session context |
| Cost tracking | Per plan and per job | Subscription or API billing |
| Source | Source-available, FSL-1.1-ALv2 | Anthropic product |
Execution model
Claude Code is interactive. You type, it acts, you respond. Plan mode writes out the steps before touching files, but the plan is a message in the session, and when the session ends it is gone unless you saved it.
Tendril makes the plan a stored, reviewable object. The CreatePlan promptware drafts it from a ticket, a GitHub issue, a jam.dev bug report, or a typed idea. The plan appears in Drafts. You read it, comment inline, and UpdatePlan rewrites it. ExpandPlan adds detail; SplitPlan breaks it into smaller plans that run separately. Only when you approve does ExecutePlan start Claude Code in a fresh worktree with the plan as its instruction. When Claude Code finishes, Verify runs your tests and lint, and the diff and the verification output appear together in Review. Approve, and CreatePr opens the pull request.
The result is exactly two human checkpoints per unit of work, at the plan and at the diff, instead of a stream of permission prompts. Claude Code still asks for its own permissions inside the worktree; Tendril adds the two gates around it.
Parallelism
You can run several Claude Code sessions by opening several terminals, and subagents can split a task inside one session. What is missing is a shared view across sessions: which are running, which finished, which failed verification, what each one cost, and whether two of them touched the same files.
Tendril runs each plan in its own git worktree on its own branch. The main branch stays clean until review. The Jobs surface streams agent output from every run. The Dashboard shows plan status, cost KPIs, a trend chart, and git activity. The Review surface is a queue of diffs waiting for a human. This is what lets one reviewer keep up with many agents. Ivy reports going from roughly 10 to more than 100 pull requests per day after adopting the workflow. See git worktrees for parallel AI agents for the mechanics and agent orchestration patterns for how to structure work so plans do not collide.
Memory and instructions
Claude Code reads CLAUDE.md files for project-level instructions. You edit them by hand as you learn what the agent gets wrong.
Tendril's promptware model works at a different level. Each stage of the lifecycle is a self-contained unit with a Program.md holding its instructions, a Memory/ directory for persistent learnings, a Tools/ directory with scoped permissions, and Logs/ with its execution history. After a run, the agent writes what it learned back to Memory/ and can revise its own Program.md. CreatePlan gets better at planning your codebase over time; ExecutePlan gets better at executing in it. CLAUDE.md keeps doing its job inside the worktree while the promptware learns at the level of the workflow.
Model choice and cost
Claude Code runs Claude models. For many teams that is the right choice, and Tendril does not change it. When you want to compare a cheaper model on routine plans, or a provider has an outage, you can switch the agent or the model per plan without changing the workflow.
Tendril records tokens and cost per plan and per job, so you can see what a class of ticket costs before you scale it up. Your traffic goes directly to the provider. Tendril does not proxy it, and nothing is trained on your code.
When Claude Code alone is the right choice
- You are one developer working on one task at a time and you are present for the whole run.
- Your tasks are exploratory and hard to plan up front.
- You are committed to Claude models and have no need to compare others.
- You do not yet have a review process for agent-written code and want to build the habit first.
Claude Code by itself is a complete tool for these cases.
When Ivy Tendril is the right choice
- You have a queue of defined tickets and want them worked in parallel while you review.
- You want a human to approve a written plan before any code changes, and a diff plus test output before any merge.
- You want cost per task, not per month.
- You want Claude Code for some plans and another agent for others, under one workflow.
- You need plans, logs, and agent memory to stay on your machine.
Install with curl -sSf https://cdn.ivy.app/install-tendril.sh | sh on macOS or Linux, or irm https://cdn.ivy.app/install-tendril.ps1 | iex on Windows.
Frequently asked questions
Does Tendril replace Claude Code?
No. Tendril orchestrates agents and Claude Code is one of them. Inside a plan, Claude Code does the editing, testing, and committing. Tendril adds planning, review, verification, parallelism, and cost tracking around it.
Do I still need CLAUDE.md?
Yes, if you use Claude Code as the agent. Claude Code reads it inside the worktree as usual. Tendril's promptware memory works at a different level: how to plan, execute, and open pull requests for your codebase.
Which is cheaper?
They are paid differently. Claude Code is billed through a Claude subscription or Anthropic API usage. Tendril's core application is free; you pay your model provider directly and Tendril shows that cost per plan and per job. Pro is $59 per user per month for team features.