Devin and Ivy Tendril both let you hand a ticket to an AI agent and get a pull request back. The difference is where the work happens and who controls it. Devin, from Cognition, is a hosted autonomous engineer: it runs in a cloud environment that Cognition operates, with the models Cognition selects, and you interact with it through a web app or chat. Ivy Tendril is a local-first desktop application: agents run on your machine, in git worktrees, using the agent and model you choose with your own API keys, and every plan and every diff passes a human before it moves on. If your code may be hosted in a vendor's cloud and you want the least involvement, Devin fits. If your code must stay on your infrastructure, or you want to choose the agent and the model and see the cost per task, Tendril is built for that.
What each tool is for
Devin
Devin is a hosted agent. Cognition provisions an environment with a shell, an editor, and a browser, connects it to your repository, and you assign tasks by chat or through the Devin web app. Devin works through the task, runs commands in its environment, and opens a pull request. You can run several Devin sessions at once and check on them. It is a closed-source product with subscription-based pricing; see Cognition's site for current plans. Its main appeal is that a team can delegate work without running anything locally.
Ivy Tendril
Ivy Tendril is a desktop application for macOS, Windows, and Linux, from Ivy Interactive AB in Stockholm, that runs the workflow the product calls a software factory. A ticket becomes a plan. A human reviews the plan. An agent executes it in an isolated git worktree. Tests and lint run. A human reviews the diff. A pull request is opened. It also runs headless with tendril --web. The core application is free and source-available under the Functional Source License, with Pro and Enterprise plans for team features, on-prem hosting, SSO, verification imports from CI, and support.
Tendril does not ship an agent or a model. It drives Claude Code, OpenAI Codex CLI, GitHub Copilot CLI, Google Gemini CLI, OpenCode, or any CLI agent, with your own keys. For the general idea, read what is a software factory.
Side by side
| Aspect | Ivy Tendril | Devin |
|---|---|---|
| Where agents run | On your machine, in a git worktree per plan | In Cognition's cloud environment |
| Where your code goes | Stays local; calls go only to your LLM provider and GitHub | Cloned into Cognition's environment |
| Model choice | Any provider you hold a key for, switchable per plan | Selected by Cognition |
| Agent choice | Claude Code, Codex CLI, Copilot CLI, Gemini CLI, OpenCode, any CLI agent | Devin |
| Human checkpoints | Plan approval and diff approval, always | Task assignment and PR review |
| Verification | Tests, lint, and diff in Review; CI import on Pro | Tests Devin runs in its environment, reported in the PR |
| Cost | Free core; you pay your provider directly; tracked per plan and job | Subscription, see Cognition's site |
| Source | Source-available, FSL-1.1-ALv2 | Closed |
| Hosting | Local desktop, headless, or on-prem for Enterprise | Hosted by Cognition |
Where code runs
This is the largest difference and it decides the question for many teams.
With Devin, your repository is cloned into an environment Cognition operates. Devin reads it, edits it, runs it, and pushes branches back. Cognition documents its security practices and many companies find them acceptable. But a third party holds a copy of your code and runs arbitrary commands against it, and that is a procurement and legal question as much as a technical one.
With Tendril, nothing leaves your machine except two kinds of traffic: prompts and responses to the LLM API you chose, and git operations with GitHub. Plans, agent memory, execution logs, and cost data are stored locally. Tendril does not proxy model traffic and there is no Tendril server that sees your code. If your company already has a data processing agreement with Anthropic, OpenAI, or Google, or routes through OpenRouter, Vercel AI Gateway, or Cloudflare, that agreement covers the only external traffic Tendril generates. We cover the reasoning in local-first AI development.
Human checkpoints
Devin's design goal is autonomy. You assign a task and Devin decides how to approach it, executes, and returns a PR. You can watch and intervene, but the default is to let it run.
Tendril's design goal is a fixed number of checkpoints. There are exactly two, and they cannot be skipped. First, the plan: the CreatePlan promptware drafts it, you read it, comment inline, expand it, split it, or approve it. No code changes until you do. Second, the diff: after the agent finishes, the Verify step runs tests, lint, and a diff summary, and the Review surface shows the change next to the verification output. Nothing reaches a pull request without your sign-off. The article on verification gates describes what runs at that step.
Which approach is right depends on how much you trust the output. For low-risk repositories, Devin's default is faster. For code that ships to customers, two mandatory reads per change is what most engineering managers want.
Parallelism and isolation
Both tools run many tasks at once. Devin does it with many cloud sessions. Tendril does it with many plans, each agent in its own git worktree on its own branch, so the main branch is untouched until review. The Dashboard shows status, cost, and git activity across plans; Jobs streams each agent's output; Review is the queue of diffs waiting for a person. Because agents run locally, they use your existing toolchain, environment variables, and private package registries with no remote environment to configure.
A laptop has finite CPU and memory, and very large parallel runs will reach that limit. Headless mode with tendril --web on a workstation or server, reachable from a phone through a Cloudflare Quick Tunnel, is how teams scale past it.
Cost control and lock-in
Devin bundles compute and model usage into its subscription. That is simple to budget, but the marginal cost of a task is not visible, and the models are whatever Cognition runs.
Tendril's core is free. You pay your model provider directly and Tendril records tokens and cost per plan and per job. When one class of ticket is expensive, you see it and can move that plan to a cheaper model or a different agent. Because Tendril is agent-agnostic, source-available, and works through ordinary git branches and GitHub pull requests, leaving it later means your repository, history, and PRs are exactly as they were. Nothing is trained on your code.
When Devin is the right choice
- You want no local setup. Team members assign work from chat and never run anything themselves.
- Your code can be hosted by a third party under Cognition's terms, and procurement has approved it.
- You prefer a single vendor to choose and operate the models.
- You want hosted compute that does not depend on anyone's laptop.
Devin is a reasonable choice for teams in that position.
When Ivy Tendril is the right choice
- Your code must stay on your infrastructure, with model traffic going only to a provider you have a contract with.
- You want to choose the agent and the model, and change them per plan.
- You want a written plan approved before execution and a verified diff approved before merge, every time.
- You want cost per task and per job.
- You want to inspect or modify the tool. The source is available under FSL-1.1-ALv2.
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. More questions are answered on the FAQ page.
Frequently asked questions
Does Tendril have a hosted option?
Tendril runs on your machine or headless with tendril --web on infrastructure you control. Enterprise adds on-prem hosting for teams. Ivy does not host your code or run your agents.
Can Tendril ingest tickets the way Devin takes tasks from chat?
Yes. Tendril ingests GitHub issues and jam.dev bug reports via webhooks, and accepts plans through the CLI, REST API, and MCP server.
Which models does Tendril use?
Whichever you configure. Documented providers include Anthropic, OpenAI, Google, OpenRouter, Vercel AI Gateway, Cloudflare, Scaleway, and NVIDIA. Both the agent and the model can be changed per plan without changing the workflow.