# Show HN: Talgent Workbench – Open-source managed agent platform you run yourself > Source: > Published: 2026-11-13 07:26:55+00:00 A managed agent platform you run yourself. Every Work in its own sandbox or on your own machine, with the model you choose. **[Live demo](https://workbench.talgent.work)** · [Website](https://talgent.work) · [Get started](#get-started) · [Documentation](#documentation) · [Releases](https://github.com/talgent-work/workbench/releases) · [Report a bug](https://github.com/talgent-work/workbench/issues/new/choose) Talgent Workbench gives every person, and every program you register, an AI Agent of their own. Start a Work, hand it files, follow what it does as it happens and collect the results. Workbench runs Claude Code natively and hosts everything around it: a sandbox per Work, sessions that resume where they stopped, credentials kept out of reach, Plugins, files, sharing and usage. All of it, including the conversations and telemetry, runs in your own Kubernetes cluster. - **The whole platform is yours.** The control plane, the Agents' conversations, their files and their telemetry live in your cluster, not only the sandboxes. Apache-2.0, so you can read and audit the code that runs your Agents. - **Any model, per Work.** Offer Claude, DeepSeek, Qwen, Kimi, GLM, MiniMax or any vendor with an Anthropic-compatible API, each with its own key. People and programs choose one when they start a Work. - **In the cloud or on your own machine.** Run a Work in a VM-isolated sandbox on your cluster, or on your own Mac, Linux or Windows computer. The conversation, Library, Usage and Plugins are the same either way. - **For people and for programs.** A web app where everyone has their own Agent, and the same public APIs for the Applications you register, each with its own Agent, Plugins and authorizations. - **The Claude Code ecosystem as it is.** Plugins come from standard Claude Code plugin catalogs, so the Skills and Agents you already use work unchanged. - **Costs and access you can trace.** Token usage by person, Work, model and Application. Every Work, file and share belongs to the person or Application that made it. Each time you start a Work, choose where it runs. Everything else stays the same. | | Cloud (hosted) | Managed Computer | |---|---|---| | **Where it runs** | In a fresh sandbox on your Workbench's Kubernetes cluster | On your own Mac, Linux or Windows machine, connected with Talgent Computer | | **Best for** | Work that should run any time, with nothing to set up and no machine left on | Work that needs the files, tools and apps already on your machine | | **Isolation** | Its own VM-isolated sandbox, cut off from your private network; credentials and model keys never enter it | You choose how much the Agent may do there, and can confine it to the operating system's sandbox on macOS and Linux | | **Setup** | None for you; the deployment provides it | Install `@talgent/computer` from npm and connect the machine from your Computers page. It needs only an outbound connection and keeps itself up to date | **Hand work to your Agent.** Describe the task, attach files, pick a repository, a model and the Plugins it needs. Steer it while it runs, stop it, and pick it up later where it left off. **Watch it work.** The conversation streams live: the Agent's thinking, the tools it calls, its plan and its progress. Share a Work with colleagues so they can follow it read-only until you revoke access. **Keep everything it made.** Reports, pages, spreadsheets and decks the Agent publishes land in your Library, next to what others have shared with you. Office documents, PDFs and images preview in the browser, and a Share Link hands a file to anyone you choose. **Extend it from the Marketplace.** Add the Plugins (Skills and specialist Agents) and Connectors your Workbench offers, and authorize your own GitHub account for the repositories a Work checks out. **See what it costs.** Token use over time, by kind and by model, and the Works that used the most. | Start a Work with what it needs | Everything your Agents made, in your Library | | Usage down to the token | Plugins and Connectors from the Marketplace | Screenshots show sample data. Register an **Application** and your own systems or product drive Agents through the same public APIs the web app uses: start Works, follow them, collect their Artifacts and read their usage. Each Application has its own Agent, Plugins and authorizations, signs in with OAuth 2.0 client credentials, labels its Works to attribute usage to its own users, and can give a Work its own MCP tools. Workbench gives every person an Agent of their own. When work outgrows one person, **Talgent Orchestra**, built on Workbench, brings those Agents together across a company's projects, so every change reaches the people it affects and finished work becomes company knowledge. - **Goals with an owner.** Every goal of a project on one board, with its owner, its priority and the Agents working on it. - **A coordinator that keeps every Agent in step.** It knows how the goals depend on each other. When one changes, the owners of the others hear about it, and their Agents adjust. - **Decisions you can trace.** Every decision the coordinator makes is on a timeline, with what it rested on, what it did and how that went. - **Knowledge that compounds.** Finished work becomes pages the whole project learns from, every statement traced to the work it came from. Say Marketing's Agent moves checkout to show prices in EUR. The coordinator sees that invoices and the Q4 report depend on it, and asks Billing and Finance. Billing needs two more days, so the launch moves by two days and every owner is told. The EU pricing rules are kept for the next time. | Every decision, with its inputs and outcomes | What the project has learned, and from where | - **Try it now** in the[live demo](https://workbench.talgent.work) : sign in with GitHub and start a Work. - **Deploy Workbench** on your Kubernetes cluster with the[self-hosting guide](https://github.com/talgent-work/workbench/blob/master/docs/self-hosting.md) . - **Connect your machine** with Talgent Computer: install`@talgent/computer` from npm and follow the Computers page of your Workbench (macOS, Linux and Windows). - **Talk to us** about Talgent at[talgent.work](https://talgent.work) . | Topic | Where | |---|---| | Deploying on Kubernetes | [Self-hosting guide](https://github.com/talgent-work/workbench/blob/master/docs/self-hosting.md) | | Every setting of every process | [Configuration reference](https://github.com/talgent-work/workbench/blob/master/docs/configuration.md) | | What an upgrade asks of an operator | [Release notes](https://github.com/talgent-work/workbench/blob/master/docs/release-notes.md) | | Processes, stores and repository layout | [Architecture](https://github.com/talgent-work/workbench/blob/master/docs/architecture.md) | | Building, testing and releases | [Development](https://github.com/talgent-work/workbench/blob/master/docs/development.md) | | Talgent Computer and its gateway | [Talgent Computer protocol](https://github.com/talgent-work/workbench/blob/master/docs/computer-protocol.md) | | Calling Workbench as an Application: tokens, APIs, identity | [Contracts repository](https://github.com/talgent-work/contracts) :`docs/application-tokens.md` ,`proto/*/public` | Report bugs and request features as [GitHub issues](https://github.com/talgent-work/workbench/issues/new/choose); the forms ask for what you did, what you expected, what happened, and the release you run. Report security vulnerabilities privately, as [SECURITY.md](https://github.com/talgent-work/workbench/blob/master/SECURITY.md) describes. The project does not accept outside pull requests. Apache License 2.0; see [LICENSE](https://github.com/talgent-work/workbench/blob/master/LICENSE).