Impatient ProgrammingIMP

Tools

The agent itself

Hooks, sandboxes and evals that put guardrails on the agent.

Some checks belong on the agent itself. A hook can stop a command before it runs, and a sandbox limits the damage when one gets through. Worktree managers let several agents work at once without stepping on each other, which is how rule 15's parallel spikes get done. This is the newest corner of the directory, and the one most likely to change by next month.

  • Start here

    Anthropic sandbox-runtime (srt)

    Restricts which files and network hosts a process can reach, with no container needed.

    The OS blocks anything outside the allow list, so an agent can run bash freely.

    Anthropic calls it a research preview.

    Swift/iOS Open source, Apache-2.05.4k starsLatest release Sep 2026Source

  • Start here

    Claude Code hooks

    Shell commands that run before or after Claude Code uses a tool, and can block it.

    A rule like no destructive commands becomes something the machine enforces on every agent run.

    Any languageWorks with Claude Code PaidPart of Claude Code

  • Start here

    promptfoo

    Runs declarative evals and red-team scans on prompts and LLM apps in CI.

    A failing eval blocks the merge when the product itself is an LLM app.

    OpenAI announced it would acquire Promptfoo in March 2026. The repo stays MIT licensed.

    InfrastructureWorks with any CI Open source, paid tier, MIT26k starsLatest release Oct 2026

  • AGENTS.md

    A plain markdown file at the repo root that tells coding agents how to work there.

    Every agent reads the same instructions, so the team keeps one file current.

    Any language Open source, MIT25k starsSource

  • Aguara

    Scans AI agent skills and MCP servers for prompt injection, exfiltration and supply-chain attacks.

    Third-party skills and MCP servers get screened before an agent loads them.

    Any language Open source, Apache-2.093 starsLatest release Sep 2026

  • Braintrust

    Hosted platform for LLM evals, traces and CI checks on prompt changes.

    A pull request fails when its prompt change makes outputs worse, before a reviewer has to notice.

    Python, JavaScript/TypeScript, Infrastructure Paid

  • Claude Squad

    Terminal app that runs several coding agents in isolated git worktrees and tmux sessions.

    The cheapest way to try parallel spikes from a terminal, with each agent on its own branch.

    Any languageWorks with Claude Code, Codex, Gemini, Aider Open source, AGPL-3.08.6k starsLatest release Aug 2026

  • Conductor

    Mac app that runs many Claude Code and Codex agents in separate git worktrees.

    Shows diffs from parallel agents side by side, so a person looks only at what matters.

    Swift/iOSWorks with Claude Code, Codex Paid

  • container-use

    MCP server that gives each coding agent its own container and git branch.

    Each agent gets a private container and branch, so a bad run never touches your machine or another agent's work.

    Dagger marks it experimental.

    Any languageWorks with Claude Code, Cursor Open source, Apache-2.04.1k starsLatest release Aug 2025

  • Cursor rules and hooks

    Project rule files and lifecycle hooks that steer and gate what Cursor's agent does.

    Rules live in the repo, and a hook can veto an action before the agent takes it.

    Any languageWorks with Cursor PaidPart of Cursor

  • DeepEval

    Pytest-style framework for testing LLM outputs with metrics like faithfulness.

    LLM quality checks run like ordinary unit tests, so a regression fails CI.

    Python Open source, paid tier, Apache-2.019k starsLatest release Sep 2026

  • Dev Containers

    A spec for describing a repo's development container in a devcontainer.json file.

    The agent works in the same locked, reproducible environment on every run.

    Any languageWorks with VS Code, GitHub Codespaces Open source, CC-BY-4.05.8k starsSource

  • E2B

    Cloud microVM sandboxes where agents run code.

    A bad command costs a disposable VM and leaves the laptop alone.

    Any language Open source, paid tier, Apache-2.014k starsLatest release Oct 2026Source

  • Fly.io Sprites

    Full Linux computers for agents with environment checkpointing and restore.

    It provides secure, disposable environments for agents to run in, with full environment checkpointing and restore.

    Any language Paid

  • Inspect (UK AISI)

    Open source framework for evaluating models and agents, from the UK AI Security Institute.

    A free, rigorous framework for turning agent evals into a merge gate.

    Python Open source, MIT2.9k starsSource

  • Nimbalyst

    Open source successor to Crystal for running agent sessions in git worktrees.

    Adds more agents and a workspace for notes and diffs to the parallel-session idea.

    Swift/iOSWorks with Claude Code, Codex Open source, MITSource

  • Sculptor

    Imbue's app that runs coding agents in isolated containers and helps merge their work.

    Containers per agent plus a pairing mode for testing an agent's changes locally.

    Swift/iOSWorks with Claude Code, Codex Open source

  • vibe-kanban

    Kanban board that dispatches coding agents to worktrees and collects their diffs.

    A task board becomes the control surface for many agents at once.

    Bloop shut down in April 2026 and the community keeps the project going.

    Any languageWorks with Claude Code, Codex Open source, Apache-2.028k starsLatest release Apr 2026

Gone quiet 2

These tools are archived or shut down, or nobody has released or committed to them in two years. Some still work fine. Read the note before you adopt one.

  • Crystal

    Desktop app for parallel Claude Code sessions in git worktrees. Now deprecated.

    Teams still running it should move to Nimbalyst, which carries the same idea forward.

    Deprecated in February 2026 in favor of Nimbalyst.

    Swift/iOSWorks with Claude Code, Codex Open source, MIT3.1k starsLast change Feb 2026

  • Daytona

    Fast sandbox environments for running agent-generated code.

    Quick start times make a fresh sandbox per task practical for parallel spikes.

    Core development moved private in June 2026 and the public repo was archived in October 2026.

    Any language Open source, paid tier, AGPL-3.072k starsLast change Jun 2026Source

Missing a tool that belongs here? Suggest it.

Every tool here was checked against its GitHub repository or its own site on 6 October 2026. Stars and dates come from GitHub on that day.