Durable agents
A run is a durable workflow, not a terminal session. When the platform restarts, the host reboots or a sandbox dies, the run resumes where it stopped. Transcripts and cost are recorded outside the sandbox.
Self-hosted agent platform · open source
Agents that keep working while you are away.
You keep the decisions that matter. Tendworks starts AI agents from triggers, runs them isolated and recoverable in sandboxes, and only calls on you when a policy says so.
$ tend status
# what runs, what waits for you
$ tend trigger fire security-review --follow
# start a workflow by hand, follow it live
$ tend chat --workspace playground
# ask an assistant about a workspaceTend: to care for something that grows. You plant the work, the platform tends it, and you come back to results and to the few decisions only you can make.
Why Tendworks
Unattended, accountable workers
Coding agents are good enough to do real work, but most setups still need a human at the keyboard: start the agent, watch it, restart it after a crash, copy the result somewhere. Tendworks takes that off your hands.
Sandboxed agents
Agents are treated as untrusted. Each runs in its own container, non-root and without privileges, with gVisor as an opt-in, and holds only the narrowly scoped credential of its workspace.
Triggered, not babysat
Trigger bindings start workflows by themselves: on a schedule, for every GitHub issue with a label, or by hand. One binding serves any number of workspaces.
Out of the loop, but in charge
Policies decide when to ask you. Notifications reach you on your phone with the priority you choose. The platform merges nothing: agents open pull requests, you decide.
Open standards at every seam
A2A, ACP, MCP, CloudEvents, OpenAPI and CNCF Serverless Workflow. Harness, model, sandbox backend and notification channel each sit behind a port.
Declarative, in Git
Workflows in YAML, agent profiles in Markdown, skills as SKILL.md, policies and trigger bindings as readable files. The web UI and the tend CLI use the same public API as your scripts.
How it works
From trigger to pull request, without you at the keyboard
Trigger
A schedule, a labeled issue or your hand starts a workflow for a workspace.
Run
A durable workflow hands the task to an agent in its own sandbox.
Ask
Where a policy says ask, the workflow waits for your decision, and you get a notification.
Deliver
The agent reports its result, such as a pull request, as an artifact. You review and merge.
Use case
Tendworks builds Tendworks
The first application is an AI-driven software factory: a value stream that turns an idea into delivered software, with humans involved only for real decisions. It is pure configuration on the platform: workflows, agent profiles, policies and skills.
And it is not a slide. A human shapes a spec and splits it into tickets; a trigger starts one sandboxed agent run per ticket, which implements it test-first, runs the checks and a code review, pushes and hands over the next ticket. The human reviews and merges.
The platform is use-case neutral: the same building blocks can run home automation or personal research routines.
- Attention at the start of the chain: what to build, not reading code.
- Evidence over trust: only what the evidence supports moves on.
- Author and reviewer are separate, ideally with a different model.
- Autonomy is earned, by the measured hit rate, not by flipping a switch.
Get started
One Linux host, one Compose stack
Tendworks is pre-1.0 and built for one operator on one Linux host. Install a release on Ubuntu 24.04 with Docker Engine, bring your own configuration, and let the agents work.