agentclientprotocol/codex-acp

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

What it does

This adapter connects Codex to editors and clients supporting ACP. It is a protocol server, not a separate language model or a complete development environment. It accepts text, images and workspace context, starts the Codex App Server, and returns replies, tool events, file changes and permission requests. Clients can select models, reasoning effort, sandbox modes and allowed user permission profiles. Native subagent sessions and some mid-turn capabilities require capability negotiation with the client. The npm package includes a compatible Codex dependency, but client launch configuration and Codex authentication remain separate prerequisites. Authentication supports ChatGPT login, API keys and negotiated custom gateways. Installation does not automatically configure every editor or establish a safe approval policy, and some permission-profile fields use an experimental schema.

License

Custom license Custom or non-standard license. Read the license file before any reuse.

More in this category

• govijr/terminal-mac
  This tool gives computer-use AI agents a separate virtual Mac. Apple Virtualization runs macOS, while a helper inside the guest captures the screen and handles input. Agents request screenshots, clicks, drags, keystrokes and status through MCP or the command line. MCP is a common communication method connecting AI tools to external functions. Requests travel as messages through a shared folder, and results return as images or structured responses. Users can view the same guest in a compatible terminal or a native virtual machine window. Getting started requires an Apple silicon Mac with macOS 14 or later, a macOS restore image and storage space. A person must create the guest account and grant Screen Recording and Accessibility permissions during initial setup. The guest does not inherit host apps or signed-in sessions, but default networking can reach the host or local network. It is an early release candidate, and official validation documentation still lists end-to-end guest installation and screen control as incomplete.
• deeplethe/sandstorm
  Sandstorm manages isolated execution environments for agent services that spend time waiting between requests. Developers supply an application template, resource limits, and sandbox creation requests. It returns sandbox state and command output through an HTTP API and provides a management console. A master keeps desired state in SQLite and distributes work to nodes. Idle environments move from running to paused, disk snapshots, and finally a state retaining only their volume. The node gateway restores an environment when traffic arrives, subject to admission and restore queue limits. It supports gVisor containers and Firecracker microVMs, with separate network namespaces for sandboxes. Setup requires Linux with cgroup v2 and backend dependencies, and Firecracker also requires a compatible virtualization environment. Snapshot placement depends on backend and host compatibility, and capacity pressure can produce refused requests rather than guaranteed immediate execution.
• rerun-build/rerun
  Rerun is a self-hostable web application for teams that want AI coworkers with separate roles and recurring work. A Foreman collects natural-language requirements and helps form a team, while a board presents coworkers and their tasks. Requests, schedules, files and connected services feed agents that can use a browser and tools and produce messages, reports, files and database records. The architecture separates a Next.js interface, a workspace engine and Supabase for authentication, storage and application data. Its memory code creates a short index of active memories for later retrieval, and its scheduler launches due work while avoiding overlapping occurrences of the same schedule. External work requires separately authorized service connections through APIs or MCP, a standard interface for connecting model tools, and model access is also configured separately. The documented self-hosting path prepares Docker, Node.js 22, Supabase and model-provider settings before creating a coworker, so self-hosting does not remove those setup and usage dependencies. Although the README promises approval before consequential actions, the inspected chat approval endpoint returns HTTP 501 as unsupported, so the claimed approval guarantee is not established by this implementation and must be checked separately.
• modelflare-ai/AnyModelBot
  AnyModelBot is a local-first assistant application for people who want persistent bots with separate personas, workspaces, models and memory. Instructions and files arrive through its desktop or local web interface and supported messaging channels, and outputs include replies, files, command results and desktop actions. A Node.js backend stores conversations in SQLite, keeps scoped Markdown memory and runs agent sessions through DeepSeek Harness with a bundled model gateway. Background tasks and scheduled rules have their own state, budget handling and a queue for questions or approval requests. Read-only, workspace and full-access presets define different execution authority, while optional desktop-action approval is distinct from file permissions. Users start with a desktop package and Modelflare authorization or an API key, then choose a model and optionally connect messaging or supported data accounts. Windows, macOS and Linux are documented, but computer control requires a desktop session, macOS control is beta, and arbitrary user-added MCP connections remain on the roadmap. Local storage does not mean offline inference: selected context, tool results and screenshots travel through Modelflare to model providers, while scheduled work also depends on the host remaining available.
• zzp1221/Research-Assistant-Agent
  Research Assistant Agent is a multi-agent research execution prototype for researchers reproducing papers, evaluating datasets and exploring candidates within a budget. Inputs are natural-language research goals, papers or attachments, repository addresses, datasets and resource constraints. A planner builds a dependency graph, meaning a plan with explicit task prerequisites, and a scheduler dispatches specialist agents with approval, retry, cancellation and budget handling. Outputs include experiment artifacts, metrics, environment manifests and links between individual claims and evidence, with separate outcomes for completed, partial, minimal-check, blocked and review-needed work. It also exposes literature search, page-based document retrieval, REST interfaces and live execution events to a React workbench. The main implementation combines a Go backend, a separate sandbox service and a Python optimizer, and the documented Compose setup requires Docker and an OpenAI-compatible model endpoint and key. Starting through that setup requires configuring service authentication and model access before opening the workbench, rather than merely opening a static page. Older root instructions say the backend is Linux-only, while the newer desktop section and desktop README describe a Windows Wails shell, native sandbox support and cross-builds, so native Windows readiness cannot be treated as settled from the overview alone. It remains a research prototype with limited production authentication and incomplete public raw verification logs. The verification document still lists claim-rubric parsing and paper-metadata extraction defects, and successful minimal runs or offline framework baselines are not full paper reproduction or real framework performance evidence.

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.