Shape agent capabilities into your workflow.

TRAIL is a benchmark dataset with 148 annotated agent execution traces and 841 errors across reasoning, execution, and planning. The traces come from software-engineering and information-retrieval tasks, and the evaluation scripts run a LiteLLM-compatible model on GAIA or SWE Bench splits before calculating scores. It is intended for studying error localization in complex agent workflows.

Binharic is a terminal coding agent with a Tech-Priest persona and support for models from OpenAI, Google, Anthropic, and Ollama. It uses built-in tools, keyword-based retrieval, predefined development workflows, and MCP extensions to handle common coding tasks. The project is still in early development, so bugs and breaking changes are expected.

Ask an agent to add temporary alternatives to a rendered surface, switch between them with a local picker, and name the visible winner in chat. Unship removes the losing options and checks that temporary artifacts are gone before release. It has no telemetry or remote session store, and is intended for local comparison rather than production experiments, analytics tests, auth, or payment flows.

Galley runs Claude Code, Codex, or GLM tasks in managed Git worktrees, records commands and diffs, and asks a supervisor to accept, retry, escalate, or stop each attempt. Repository profiles define checks, scope, secrets policy, and optional pull-request handoff, while evidence remains on disk for later inspection. Task YAML is trusted execution input, so Galley should be used only with task authors and repositories you trust.

Awesome AI Agents is a categorized list of frameworks, tools, platforms, starter kits, examples, and learning resources for building autonomous or semi-autonomous agents. It covers ideas such as planning, memory, tool use, and multi-agent coordination, so it is useful for comparing options rather than running an agent itself.

LionClaw wraps existing agent CLIs such as Codex and OpenCode in a project-local control plane. It keeps runtime profiles, durable sessions, skills, channels, scheduled jobs, private state, and audit records under a `.lionclaw/` directory while preserving the underlying agent’s own interface. Use it when you want swappable runtimes with explicit control over workspace access and credentials.

Agent Memory System creates repository-local Markdown and JSON artifacts covering project structure, commands, interfaces, security context, dependency graphs, worklogs, and handoffs. Coding agents can read the generated index and query the graph before editing shared code, then record checkpoints for the next session. It is not a chatbot memory service or agent runtime, and its benchmark results are maintainer-run early measurements rather than independent validation.

Multi-Agent Workflow Kit creates a separate Git worktree, branch, and directory for each coding agent, then places them in split panes within one tmux session for supervision and messaging. Agents are configured in .agents/agents.yaml and managed with maw commands for provisioning, syncing, navigation, and communication. The project is still an early proof of concept and requires Git 2.5+, tmux 3.2+, yq, and uvx.

agent-terminal lets an AI agent launch interactive CLI applications, capture their screen as ASCII text, and send keyboard input without a display. It can be used as an npm library or exposed through an MCP server for CLI testing, monitoring, and build automation. Built-in safeguards include command allowlists, workspace path validation, environment sanitization, input checks, and session limits.

Praman turns a business-process description into planned, generated, and repaired Playwright tests for SAP GUI, UI5, WebGUI, and Fiori applications. For UI5-based interfaces, it queries controls through the runtime registry rather than relying only on fragile DOM selectors, while also offering typed fixtures for manual test authoring. It requires Node.js 22 and a compatible Playwright release.

Agentic Engineering Framework wraps CLI coding agents with task gates, human approval for destructive actions, persistent session memory, dependency mapping, and audit trails. It coordinates and governs tools such as Claude Code, Cursor, or Aider rather than running the model itself. The framework is intended for long-lived repositories that need traceable commits and controlled agent autonomy, and it is still alpha software.

CLITrigger combines project docs, planning, live terminals, autonomous tasks, and Git review in one command center. Claude, Antigravity, or Codex jobs can run in parallel or on schedules, each inside an isolated worktree, while completed diffs flow into a central accept, reject, or merge queue. It is available as a desktop app or a Node-based service for developers managing many AI coding tasks.

Grite stores issues and task history in Git, so they branch, merge, and sync alongside the code through ordinary fetch and push operations. Agents can claim work with distributed locks, record progress and architectural memory across sessions, and model dependencies between tasks. Its deterministic merge model is designed to preserve concurrent updates without requiring an external tracker, account, database, or network service beyond Git itself.

Better OpenCode MCP exposes the OpenCode CLI to Claude Code and other MCP clients, enabling selectable models and non-blocking long-running tasks. Calls return a task ID that can be monitored or supplied with additional input, while process pooling, timeouts, and persisted task state help manage concurrency and recovery. It requires Node.js and a working OpenCode installation.

pi-boss breaks a coding task into parallel subtasks and launches each worker in its own visible tmux pane. Its spawn tool creates, lists, highlights, and stops panes, while the required pi-room integration provides the peek and steer controls used to monitor and guide agents. The setup also requires pi and tmux.

Not Human Search indexes websites using signals such as llms.txt, OpenAPI descriptions, structured APIs, and MCP endpoints, then ranks them for agent-oriented use. People can search through the site or REST API, while compatible agents can use its MCP server to find services, inspect site details, and check a URL score. It is aimed at developers building machine-readable web workflows.

Copilot Brag Sheet records files changed, pull requests, and git activity during Copilot CLI sessions, then stores the data locally for later review and work-log generation. It is designed to turn forgotten tasks into evidence-backed impact statements for performance reviews. Automatic session tracking currently depends on Copilot CLI, while other MCP hosts can use the tools only through explicit manual capture.

cowork-to-code-bridge connects Claude Cowork or a browser chat to Claude Code on a macOS, Linux, or WSL2 machine. It forwards approved tasks or scripts without opening network ports or using sudo, streams results back, and caches idempotent retries so work is not run twice.