ag-ui-protocol/ag-ui

This page introduces an external open-source repository. It is not an HDATF product.

What it does

AG-UI is an event-based protocol and SDK collection for developers connecting agent backends to user-facing applications, rather than a complete assistant product or agent reasoning engine. It addresses the repeated need to translate each framework's messages, tool activity, execution status, and shared state into application-specific interfaces. RunAgentInput carries the request context, including conversation messages, thread and run identifiers, state, and available tools, while the response is a stream of typed events. These events describe text fragments, tool calls, execution start or finish or failure, and state updates so a frontend can show progress and participate in the workflow. Full state snapshots replace the current state, while JSON Patch deltas describe incremental changes that clients must apply in order. The protocol is transport independent, but the inspected TypeScript HttpAgent makes JSON POST requests, accepts a server-sent event stream, validates outgoing input, and supports aborting the request. Starting requires a compatible agent endpoint or framework adapter, a client that consumes the events, and application-defined state and tool behavior. Authentication, business permissions, durable storage, conflict handling, and recovery policies still need application design, and the README's approximate event count should not be treated as a fixed limit of the current source.

License

MIT Permissive. Commercial use and changes are allowed if the copyright notice is kept.

More in this category

• backnotprop/plannotator
  Plannotator is a local browser-based review product for people supervising coding agents and reviewing their plans, documents, messages, HTML artifacts, and code changes. It addresses feedback scattered across terminal conversations by letting reviewers attach comments to the material being reviewed and return them to the agent. Inputs include local files, folders, URLs, repository diffs, and GitHub or GitLab review links, while outputs include annotations, approval or revision decisions, and feedback for the originating session. Plugins, hooks, and commands connect supported agents to a local review server, and the Claude Code configuration includes interception of ExitPlanMode permission requests. The separate Inbox surface groups questions, files, and guided reviews from multiple agents and can receive replies without keeping the original agent turn waiting. Dedicated plugins provide session wake-up for selected agents, while other agents can connect through MCP, the interface for calling external tools, so delivery behavior depends on the connector. Starting requires the Plannotator binary, a browser, and the integration appropriate to the agent, with authenticated provider CLIs needed for remote pull request review. Review data stays local by default, but URL fetching, configured AI providers, sharing, and hosted Workspaces have separate network boundaries, and remote mode exposes an unauthenticated listener that must be protected.
• libracoder/kiri
  Kiri is a Rust container runtime that accepts Docker images and commands and runs containers locally. It is intended for developers who want a different engine while retaining the docker CLI, Compose and buildx. Requests use the Docker Engine API, the protocol used to control containers, and return execution state, logs and process output. It uses a Linux virtual machine on macOS, the WSL 2 Linux environment on Windows and the host kernel on Linux. Its internal agent creates containers with operating system namespaces and cgroups to separate processes and resources, without putting runc or dockerd in that execution path. It also provides a management window for images, volumes and networks, single-node Kubernetes and separate Linux working environments. Initial setup creates the kiri Docker context, and images absent locally require access to their image registry. Builds use a managed BuildKit container, so complete Docker build equivalence should not be assumed, and the documentation still lists a buildx host-gateway limitation.
• serveroshq/panel
  ServerOS panel is a web control panel for operating Linux servers you already own. Operators supply machine enrollment details, Git repositories or deployment templates and workflow settings to manage deployments, monitoring and backups. Each managed machine needs a separately connected serverosd daemon; installing the panel does not automatically enroll servers. It covers CPU, memory and disk status, service logs, scheduled workflows and team activity. The inspected API routes expose tool calls behind API-key authentication, and the project documents MCP and API connections for external tools. Self-hosting requires a Docker-capable Linux server, a domain and, in the standard setup, ports 80, 443 and 8443 plus wildcard DNS. Email delivery and private GitHub repositories require separate SMTP and GitHub App configuration. The community edition built from this repository differs from the official image that includes commercial extension code. Panel code is AGPL-3.0, while paid extensions and trademarks have separate terms, and default anonymous installation telemetry should be reviewed.
• JenkinWoo/sheeridVerify
  sheeridVerify is a Node.js workbench for learning and authorized testing of SheerID verification workflows. Its intended users are educators, researchers, and testers with permission to handle the submitted information. A job takes a verification link, a name, a birth date, and configured school and document sources, then produces queue progress, logs, and a result reflecting the remote state. It queries the current verification step before choosing browser form interaction or document upload. Documents are read as original PNG, JPEG, or PDF bytes rather than generated or altered, and interactive human challenges remain with the user. The web workbench controls concurrency and allows cancellation, resumption, and deletion, but its queue and history are held in memory and disappear on service restart. When a submission response is uncertain, the workflow queries the remote state instead of immediately repeating the write. It requires Node.js 22.15 or newer, Camoufox for the browser branch, and access to SheerID and any configured document or proxy service. Real personal information can reach those services, and deleting a local task does not revoke a remote submission or erase disk logs. Constructing an email string does not prove ownership or student status, and the project disclaims affiliation with SheerID and any guarantee of successful verification.
• render-oss/render-mcp-server
  Render MCP Server is Render's official bridge between an AI application and resources hosted on Render. It uses MCP, a protocol through which an AI application discovers tools and passes structured parameters to them. Inputs include workspace and resource identifiers, log filters, service settings or SQL, and outputs include resource details, operational records and query results. Tools cover service and database creation, deployment history and triggering, logs, metrics and environment-variable replacement. The Postgres query tool wraps queries in a read-only transaction, but other tools can change infrastructure or initiate deployments. Clients can connect to the hosted endpoint or run a local executable or container with appropriate credentials. Database queries use the external database connection and remain subject to its IP allowlist. This is not complete Render administration, and sensitive information can still reach the AI context even though the server attempts to reduce exposure.

Only repositories in the ranked Trendshift lists are included, and the lists are used only to find candidates. We do not copy their ranks. Descriptions, licenses and star counts come from each GitHub repository. The notes are our own reading. We have not tested these projects, and a place on a trending list does not prove quality.