Find the open-source solution that fits.

Keen Code is a terminal coding agent for editing, debugging, planning, refactoring, and reviewing software with several model providers. Its deliberately small toolset covers file operations, search, and shell commands, while sessions, skills, and configurable thinking support longer work. Cross-turn context is kept lean through summaries focused mainly on changed files and failed commands.

Skillreaper inspects local configurations and session transcripts to show which skills, MCP servers, and agents load without firing, along with estimated context cost. It starts read-only, assigns REAP, MUTE, KEEP, or REVIEW verdicts, and makes pruning reversible through quarantined files and manifests. The tool is local with no telemetry or uploads, but platforms without readable history remain conservatively marked REVIEW.

RoleCraft is a zero-dependency CLI for installing agent skills from local folders, Git repositories, SSH URLs, or npm packages. It can target many compatible agents, install declared MCP servers, keep a lockfile, check updates, verify content hashes, and reinstall in CI. Every installation receives a local static security scan for prompt injection, command injection, credential harvesting, and sensitive file access, with dangerous packages blocked unless the user explicitly overrides the gate.

Agent AFK provides an Apache-licensed harness whose prompts, gates, routing, skills, traces, and provider settings are editable. It supports terminal chat, headless background runs, Telegram notifications, orchestration commands, and migration of plugins or MCP servers from Claude Code or Codex. Node 22 is required, and optional remote notifications need their own credentials.

Clave is a macOS desktop workspace for running Claude Code, Codex, Antigravity CLI, and regular terminals in split or grid layouts. It adds Git history and diffs, local and remote file access, and SSH-based sessions without a cloud account. An optional daily anonymous usage ping can be disabled in settings.

Place the scanner between untrusted content, model output, and tool execution, or wrap a line-delimited stdio MCP server with its proxy. It detects prompt injection, credential disclosure, exfiltration, and destructive actions before forwarding a call, redacts secrets, and returns structured reasons for enforcement. The scanner makes no network calls and ships as a Rust CLI with Node and Python wrappers.

pi-reflect compares recent conversations and reference material with a target Markdown file, then applies small edits to files such as AGENTS.md, MEMORY.md, or SOUL.md. It backs up changes, can commit them to Git, and tracks correction trends and repeated rule failures. Each run requires pi and a configured LLM API key.

smelt combines an AI coding agent, a Vim-style editor, and its own terminal renderer in a compact CLI. Lua plugins can add keymaps, commands, tools, and modes, while provider support covers subscriptions, OpenAI-compatible APIs, and local Ollama models. The project is under active refactoring, so the latest prerelease is recommended over the older stable build.

The plugin turns an idea into numbered acceptance criteria, decomposes the work into a dependency-ordered task graph, and runs tasks in parallel isolated Git worktrees with test-driven implementation. Code review and requirement-level verification gate atomic merges, while disk-backed state supports recovery after crashes or context resets. Remote pushes still require explicit user approval.

AgentDiff captures AI-assisted edits across Claude Code, Cursor, Copilot, Codex, Windsurf, OpenCode, and Gemini, signs line-level attribution with Ed25519, and stores the evidence in Git. Its blame, diff, context, and report commands show who changed what. Pushed traces can expose short prompt excerpts to repository readers, so sensitive teams may need to disable prompt capture.

ToolRAG registers tools from MCP servers, embeds their descriptions, and uses semantic retrieval to select the functions relevant to a user query. The selected definitions can be passed in an OpenAI-compatible format and executed against the corresponding servers, while LibSQL stores the tool index and embeddings. It is aimed at assistants that need access to many APIs without sending every tool definition on every request.

AgentManager inventories AI development CLIs on macOS, Linux, and Windows, checks their versions in parallel, and installs or updates them through package managers such as npm, PyPI, Homebrew, and native installers. It offers both CLI and interactive terminal interfaces, ships an offline-first catalog, and can expose management through a tray helper, REST API, or gRPC.

EchoCoding connects sound effects, text-to-speech, speech recognition, voice prompts, and ambient audio to Claude Code, Cursor, Windsurf, Codex CLI, and Gemini CLI through hooks, MCP, and command-line tools. Its installer starts a local daemon without requiring model downloads or API-key setup, while microphone features still depend on local audio permissions.

AGX brings tickets, repositories, agent sessions, implementation work, and pull-request review into a local CLI and dashboard with durable checkpoints for resuming later. It supports Claude, Codex, Gemini, and Ollama with human approval gates before irreversible actions; the CLI needs Node 22.16 or newer and at least one provider CLI, while anonymous telemetry is enabled by default but can be disabled.

agent-runbook lets skill authors describe multi-agent workflows in YAML and compile them into SKILL.md files with checkpoint support. Steps declare typed inputs and JSON Schema outputs, while the compiler checks references, dependency closure, and cycles before producing the artifact. Runbooks can dispatch agents, run scripts, branch, loop, execute parallel work, and resume from a checkpoint; the project notes that its API is still changing.

Puppeteer Real Browser MCP Server starts and owns a local Chrome process for an MCP client, exposing navigation, content reading, element lookup, clicks, typing, waits, and scrolling. It defaults to a visible browser, supports headless mode, proxies, custom Chrome paths, and dedicated profiles, and closes only the process it launched. The workflow requires reading page content before interaction, allows one browser session, and does not provide a real CAPTCHA solver.

AgentPack performs a local preflight before an agent edits code. Given a task, it ranks likely files, tests, rules, skills, commands, and warnings, then renders a compact context pack with reasons and receipts for what was included or skipped. It uses local analysis and reusable caches rather than cloud indexing or model calls, and its benchmark measures file-selection overlap—not guaranteed task success—so source inspection and tests still matter.

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.