Free agents, skills and workflows
Section

Free agents, skills and workflows

Everything here is free to install. Each item says how well it has been tested, and none of the labels are rounded up.

Free to installlabels from recorded test runs

37 free items · 6 workflows · 52 individual skills, agents and hooks inside the Foundry plugins

Add the AI Foundry marketplace once

claude plugin marketplace add ai-foundry-byu/foundry-marketplace

Then install any Foundry plugin by name, or type /plugin inside Claude Code to browse. The Foundry's own item pages are at its marketplace site.

What the labels mean
  • verified Passed a live end-to-end run of its main path in a real Claude Code session.
  • beta Partly live-tested or unit-tested, with known gaps.
  • new Little or no live testing yet. Try it on something low stakes first.
  • unrated Not part of the Foundry's recorded test runs. Its README states its own limits.

Plugins

Install a whole set in one command. Updates come with it.

  1. betaStarter set · AI FoundryMIT

    foundry-sdlc

    The 8-stage 0-to-1 workflow: scaffold, requirements, PRD and stories, workplan, parallel lanes, QA, ship, feedback. Templates, a PRD-to-stories compiler, a stage detector, and two independent judge agents (prd-critic, story-verifier).

    Labeled by its weakest item. 10 verified, 1 beta.

    Terminal

    claude plugin marketplace add ai-foundry-byu/foundry-marketplace && claude plugin install foundry-sdlc@ai-foundry
    Known limits
    • Projects scaffolded before this release keep an old tools/gen_stories.py; run gen_stories.py --refresh-tools if the stage readout sticks at 4.
    • Parallel lanes are protected only as subagents with worktree isolation.
    • On smaller models, call a skill by name. Triggering is tuned for Sonnet and Opus.
    What is inside: 2 agents, 1 hook, 8 skills

    Source · Item pages

  2. betaStarter set · AI FoundryMIT

    foundry-guardrails

    Hooks that make the gates real: block destructive commands and secret leaks, require tests before done and before push or deploy, keep worktree sessions in their worktree, flag stale whole-file writes, and keep working until a workplan milestone is checked.

    Labeled by its weakest item. 6 verified, 1 beta.

    Terminal

    claude plugin marketplace add ai-foundry-byu/foundry-marketplace && claude plugin install foundry-guardrails@ai-foundry
    Known limits
    • A seatbelt, not a vault. The hooks match file names and command shapes, so an unusual command such as an archive round trip can still get past them.
    • On smaller models, call a skill by name. Triggering is tuned for Sonnet and Opus.
    What is inside: 6 hooks, 1 skill

    Source · Item pages

  3. betaStarter set · AI FoundryMIT

    foundry-qa

    Web QA: a definition-of-done gate (deployed SHA plus a Playwright journey on desktop, phone and tablet), blast-radius regression checks, axe-core accessibility, screenshot baselines, and web-qa, persona-tester and explorer agents.

    Labeled by its weakest item. 6 verified, 3 beta.

    Terminal

    claude plugin marketplace add ai-foundry-byu/foundry-marketplace && claude plugin install foundry-qa@ai-foundry
    Known limits
    • Signed-in QA needs a test-only sign-in route that is off in production.
    • The full stage 6 check is heavy: each verifier and persona runs its own Chromium.
    • On smaller models, call a skill by name. Triggering is tuned for Sonnet and Opus.
    What is inside: 3 agents, 6 skills

    Source · Item pages

  4. betaAI FoundryMIT

    foundry-design

    Design gates: a subtract-only design critic, a UI spec writer with a full state matrix and checkable acceptance criteria, a design-system template, a read-only ui-reviewer agent, and a deterministic design gate script.

    Labeled by its weakest item. 2 verified, 2 beta.

    Terminal

    claude plugin marketplace add ai-foundry-byu/foundry-marketplace && claude plugin install foundry-design@ai-foundry
    Known limits
    • ui-reviewer and design-system are beta: their latest fixes have not been re-run live.
    • On smaller models, call a skill by name. Triggering is tuned for Sonnet and Opus.
    What is inside: 1 agent, 3 skills

    Source · Item pages

  5. newAI FoundryMIT

    foundry-devices

    Native and desktop QA agents for iOS Simulator, Android, Windows and macOS, plus a device-matrix skill that runs them as parallel lanes and merges one PASS/FAIL/INCONCLUSIVE matrix.

    Labeled by its weakest item. 1 verified, 6 beta, 1 new.

    Terminal

    claude plugin marketplace add ai-foundry-byu/foundry-marketplace && claude plugin install foundry-devices@ai-foundry
    Known limits
    • The Android, iOS Simulator, macOS and Windows commands pass on CI runners but have not run on a student's own app yet.
    • macOS automation needs the Accessibility permission for your terminal.
    • On smaller models, call a skill by name. Triggering is tuned for Sonnet and Opus.
    What is inside: 4 agents, 4 skills

    Source · Item pages

  6. newAI FoundryMIT

    foundry-iphone-mirror

    Drive a real iPhone from a Mac through macOS iPhone Mirroring: screenshot, read, tap, type, swipe, verify. A skill, an iphone-driver agent, an MCP server and a CLI with a doctor.

    Labeled by its weakest item. All 2 items new.

    Terminal

    claude plugin marketplace add ai-foundry-byu/foundry-marketplace && claude plugin install foundry-iphone-mirror@ai-foundry
    Known limits
    • It has never driven a real phone. Treat it as experimental.
    • On smaller models, call a skill by name. Triggering is tuned for Sonnet and Opus.
    What is inside: 1 agent, 1 skill

    Source · Item pages

  7. betaAI FoundryMIT

    foundry-evals

    Measure agents and prompts: scaffold an eval suite with oracles, code graders, a binary judge and negative controls, run and compare it, and hill-climb one change at a time. Ships evalkit.py (stdlib only).

    Labeled by its weakest item. 3 verified, 1 beta.

    Terminal

    claude plugin marketplace add ai-foundry-byu/foundry-marketplace && claude plugin install foundry-evals@ai-foundry
    Known limits
    • A suite's score is not held-out accuracy unless you keep a test split and score it separately.
    • On smaller models, call a skill by name. Triggering is tuned for Sonnet and Opus.
    What is inside: 1 agent, 3 skills

    Source · Item pages

  8. betaAI FoundryMIT

    foundry-builder-kit

    Day-one build skills: add auth, a 10k-user launch audit, secrets hygiene with a commit guard, cheap structured LLM routing, scheduled headless agents, and driving interactive CLIs through tmux.

    Labeled by its weakest item. 4 verified, 3 beta.

    Terminal

    claude plugin marketplace add ai-foundry-byu/foundry-marketplace && claude plugin install foundry-builder-kit@ai-foundry
    Known limits
    • add-auth, scheduled-agents and tmux-agents produce code that is unit-tested but not run end to end. Build and run it yourself before you trust it.
    • On smaller models, call a skill by name. Triggering is tuned for Sonnet and Opus.
    What is inside: 1 hook, 6 skills

    Source · Item pages

  9. unratedJDMIT

    claude-bug-squash

    Report a bug and it triages, fixes in an isolated git worktree, proves the bug goes red to green, and gates every fix behind blast-radius caps, a never-touch list and a three-seat review panel.

    Not part of the Foundry's recorded test runs. Its README states its own limits.

    Terminal

    claude plugin marketplace add jddavenportOpen/claude-bug-squash && claude plugin install claude-bug-squash@claude-bug-squash
    Known limits
    • An autonomous merge needs a repro command that fails while the bug is present and passes once it is fixed. Without one, fixes cannot auto-merge.
    • At most one auto-merge a day by default.

    Source

  10. betaJDMIT

    recruit (recruit-copilot)

    Merge the resumes you already have into one experience bank, score open roles against goals you set, tailor a resume per posting, and check that a machine can still read it. It never applies for you.

    Labeled beta in the Foundry marketplace's own recorded runs.

    Terminal

    claude plugin marketplace add jddavenportOpen/recruit-copilot && claude plugin install recruit@recruit-copilot
    Known limits
    • The round-trip check proves the text is present and readable in order. It cannot prove every commercial ATS parses it, so never call it ATS-verified.
    • The built-in PDF reader handles common resume PDFs, not every PDF.

    Source · Write-up

  11. unratedJDMIT

    Clawdacademy (ai-fluency-trainer)

    Scores seven judgment behaviors in your real Claude Code sessions, names your weakest habit and coaches in the flow of work.

    Not part of the Foundry's recorded test runs. Its README states its own limits.

    Its one-line installer runs a script from a third domain. Read the repo's README and install script before you run it.

    Known limits
    • Sync is opt-in and needs a token. Read the data section of the README before you turn it on.

    Source

Skills

One folder your agent loads when the task matches.

  1. unratedJDMIT

    contact-list-builder

    Builds a verified list of public business contacts for customer interviews or prospecting. Parallel research lanes, then a script re-downloads every source page to prove each email and phone is there.

    Not part of the Foundry's recorded test runs. Its README states its own limits.

    Terminal

    git clone https://github.com/jddavenportOpen/contact-list-builder ~/.claude/skills/contact-list-builder
    Known limits
    • Public pages only, and guessed emails are never shown as real.
    • Needs Python 3. pdftotext is optional, for contacts that appear only in PDFs.

    Source

  2. unratedJDMIT

    instinct-handoff

    Routing rules, a Claude Code skill and an AGENTS.md for handing work from your agent system to Instinct, an outside autonomous agent.

    Not part of the Foundry's recorded test runs. Its README states its own limits.

    Terminal

    git clone https://github.com/jddavenportOpen/instinct-standing-doc && mkdir -p ~/.claude/skills && cp -r instinct-standing-doc/skills/instinct-handoff ~/.claude/skills/
    Known limits
    • Unofficial and not affiliated with Instinct.
    • Useful only if you use Instinct.

    Source

Agents

Workers you hand a job to: research, phone calls, judging, orchestration.

  1. unratedJDMIT

    deep-research-agent

    An iterative research agent: a knowledge-gap loop over nine parallel sources, a devil's-advocate pass, and per-claim three-vote verification.

    Not part of the Foundry's recorded test runs. Its README states its own limits.

    Terminal

    git clone https://github.com/jddavenportOpen/deep-research-agent.git && cd deep-research-agent && pip install -r requirements.txt
    Known limits
    • Live runs need your own API keys in .env.
    • Its tests run offline with stubs; a live run is what costs money.

    Source · Write-up

  2. unratedJDMIT

    ai-phone-desk

    Gives your agents a phone. Nothing dials until a human approves the exact request.

    Not part of the Foundry's recorded test runs. Its README states its own limits.

    Terminal

    git clone https://github.com/jddavenportOpen/ai-phone-desk && cd ai-phone-desk && pip install -e .
    Known limits
    • Rehearsal is text, not voice.
    • Calls cost money (Vapi bills per minute).
    • US and Canada numbers only.

    Source

  3. unratedJDMIT

    orchestra-agents

    Multi-agent orchestration and an eval harness on Claude: fleet fan-out, a durable message bus, a circuit breaker, best-of-N with an adversarial judge, and a definition-of-done gate.

    Not part of the Foundry's recorded test runs. Its README states its own limits.

    Terminal

    git clone https://github.com/jddavenportOpen/orchestra-agents && cd orchestra-agents && pip install -e ".[dev]"
    Known limits
    • The tests run offline; live runs need your own Anthropic API key.

    Source

  4. unratedJDMIT

    mcp-judge

    A calibrated multi-persona LLM judge exposed as an MCP server, with an eval harness that shows the calibration.

    Not part of the Foundry's recorded test runs. Its README states its own limits.

    Terminal

    git clone https://github.com/jddavenportOpen/mcp-judge && cd mcp-judge && pip install -e .
    Known limits
    • Scoring and the eval harness need your own Anthropic API key.

    Source

Hooks

Rules that hold on every tool call, not just when the agent remembers.

  1. unratedJDMIT

    ethos-gate

    A PreToolUse safety gate that blocks irreversible or outward-facing actions (sends, spends, deploys) unless a human approved that exact content.

    Not part of the Foundry's recorded test runs. Its README states its own limits.

    Terminal

    git clone https://github.com/jddavenportOpen/ethos-gate && cd ethos-gate && pip install -e .
    Known limits
    • Install from source. The README's pip install ethos-gate line does not work yet: the package is not on PyPI.

    Source

Harnesses and tools

The machinery around the agents: orgs, fleets, governors, traces.

  1. unratedJDAGPL-3.0 · share changes if you run it as a service

    Clawdling

    Your Claude Code as a whole org: an agent for each project on one map, showing who is working and who needs you, with a receipt when one says done. Local, on your own Claude plan.

    Not part of the Foundry's recorded test runs. Its README states its own limits.

    Paste into Claude Code

    Set up Clawdling for me from https://github.com/jddavenportOpen/clawdling and follow SETUP.md
    Known limits
    • Needs a paid Claude plan (Pro, Max, Team or Enterprise). On Pro, run one or two agents at a time.
    • AGPL-3.0: if you run a modified version as a service, you must share your changes.

    Source · Write-up

  2. unratedJDMIT

    clawd-agent-os

    Tiered Claude agent hierarchies: worktree-isolated fleet spawning, a durable agent bus that never silently returns empty, and one registry for model tiers.

    Not part of the Foundry's recorded test runs. Its README states its own limits.

    Terminal

    git clone https://github.com/jddavenportOpen/clawd-agent-os && cd clawd-agent-os && pip install -e ".[dev]"
    Known limits
    • Needs the Claude Code CLI on your PATH.
    • The durable bus needs a Postgres or Supabase table you provide.

    Source

  3. unratedJDMIT

    warden

    A governor for autonomous agents: autonomy is admitted by how reversible an action is, not by how confident the agent sounds. Zero dependencies.

    Not part of the Foundry's recorded test runs. Its README states its own limits.

    Terminal

    git clone https://github.com/jddavenportOpen/warden && cd warden && pip install -e .
    Known limits
    • It checks that an undo exists, not that the undo works.
    • It governs actions, not prompts. It is not a jailbreak defense.

    Source

  4. unratedJDMIT

    fleetwright

    A self-hosted agent operations OS: a persistent chief of staff that commands a fleet of agents, with a bridge and a web cockpit.

    Not part of the Foundry's recorded test runs. Its README states its own limits.

    Terminal

    git clone https://github.com/jddavenportOpen/fleetwright.git && cd fleetwright && bash scripts/bootstrap.sh
    Known limits
    • The bridge needs access to your own claude CLI.
    • The cockpit is a Node app you run locally.

    Source

  5. unratedJDMIT

    ccstitch

    A local trace store for Claude Code's OpenTelemetry spans that repairs orphaned headless turns and nests child sessions under the tool call that launched them.

    Not part of the Foundry's recorded test runs. Its README states its own limits.

    Terminal

    pipx install git+https://github.com/jddavenportOpen/ccstitch && ccstitch install --dry-run
    Known limits
    • Depends on a Claude Code telemetry beta; span names can change in any release.
    • Single machine, single user, no authentication beyond the local bind.
    • Costs it shows are list-price estimates.

    Source

  6. unratedJDMIT

    harnessview

    Inspect and visualize your agent harness (agents, skills, hooks, MCP servers) and its run history, and spot broken config.

    Not part of the Foundry's recorded test runs. Its README states its own limits.

    Terminal

    git clone https://github.com/jddavenportOpen/harnessview && cd harnessview && npm install && npm run dev
    Known limits
    • Run it from a clone. The README's npx harnessview line does not work yet: the package is not on npm.

    Source · Write-up

Templates and specs

Starting points your agent builds out in your own stack.

  1. unratedJDMIT

    context-kit

    Four Markdown templates and five Claude Code skills that give an agent permanent personal context: a wiki, mental models, voice and protocols.

    Not part of the Foundry's recorded test runs. Its README states its own limits.

    Terminal

    git clone https://github.com/jddavenportOpen/context-kit.git && cd context-kit && bash scripts/install.sh
    Known limits
    • The templates are yours to fill. What you put in them is what your agent will know.

    Source

  2. unratedJDMIT

    claude-deploy-kit

    Enterprise controls for deploying Claude agents: a policy chokepoint, a destructive-command deny hook, an eval gate, a redacting audit log and pluggable secrets.

    Not part of the Foundry's recorded test runs. Its README states its own limits.

    Terminal

    git clone https://github.com/jddavenportOpen/claude-deploy-kit && cd claude-deploy-kit && pip install -e ".[dev]"
    Known limits
    • Identity, secrets backends and data residency are decisions the code leaves to you; the architecture doc covers them.

    Source

  3. unratedAI FoundryMIT

    app-feedback-loop

    A build spec for an in-app feedback loop: one click captures a screenshot and context, an agent fixes the bug with that context loaded, and the fix commit thanks the reporter.

    Not part of the Foundry's recorded test runs. Its README states its own limits.

    Paste into Claude Code

    Read https://github.com/ai-foundry-byu/app-feedback-loop/blob/main/SPEC.md and implement this feedback loop natively in my stack. When you believe you are done, run the checks in VERIFY.md and show me the results.
    Known limits
    • It is a spec, not a library. Your agent builds it in your stack, so review what it builds.

    Source

  4. unratedAI FoundryNo license file yet · ask before reuse

    agent-team-kit

    A working setup for running many Claude Code agents on one app without collisions: port assignment, a schema lock, git guards and gates, plus mockup and bug-fix skills.

    Not part of the Foundry's recorded test runs. Its README states its own limits.

    The repo has no license file yet, so we do not list an install line. Read it, and ask before you reuse it.

    Known limits
    • No license file yet.

    Source

Guides

Learn by running it.

  1. unratedJDMIT

    agent-evals-starter

    Learn agent evals by running them: a support agent, 24 tasks, code graders, an LLM judge, and do-nothing and pushover controls. Runs offline.

    Not part of the Foundry's recorded test runs. Its README states its own limits.

    Terminal

    git clone https://github.com/jddavenportOpen/agent-evals-starter.git && cd agent-evals-starter && pip install -r requirements.txt
    Known limits
    • The judge labels were drafted with AI for teaching. Real judge labels must come from a person.
    • 24 tasks catch big differences, not small ones.

    Source · Write-up

  2. unratedJDMIT

    mastery-engine

    Generates a curriculum from source material on disk, with mastery gates and claims checked against the source.

    Not part of the Foundry's recorded test runs. Its README states its own limits.

    Terminal

    git clone https://github.com/jddavenportOpen/mastery-engine && cd mastery-engine && pip install -e .
    Known limits
    • Version 0.1.0. No learner has yet gone from a generated curriculum to a cleared readiness gate over real calendar time.

    Source

  3. unratedAI Foundrysee repo

    The Software Design Canon

    A curated, link-checked index of the free primary sources on software design and architecture. No pirated books.

    Not part of the Foundry's recorded test runs. Its README states its own limits.

    Nothing to install. Read it, or point your agent at it.

    Known limits
    • The repo records who holds the rights to each source; check before you copy one.

    Source

Workflows

Several of the items above chained into one job, with a check at the end.

  1. Workflow

    0 to 1 product build

    Go from an idea to a shipped app with agents, one stage at a time, with a judge at each gate. Built on foundry-sdlc.

    Read the steps · 2 min

  2. Workflow

    Parallel lanes with worktrees

    Run several coding agents on one repo at once without them overwriting each other. One git worktree per lane, and a guard that keeps each lane in its own tree.

    Read the steps · 2 min

  3. Workflow

    Ship with a done gate

    Never call a feature done until production serves the commit you meant and a browser has walked the user's path on it. Built on foundry-qa.

    Read the steps · 1 min

  4. Workflow

    Bug report to merged fix

    Turn a user's bug report into a failing test, a fix in its own worktree, and a merge that is gated on proof. Root cause first, not a patch on the symptom.

    Read the steps · 2 min

  5. Workflow

    Scheduled agents and watchers

    Run an agent on a schedule, or until a condition fires, without it running forever or failing in silence. Built on foundry-builder-kit.

    Read the steps · 2 min

  6. Workflow

    Eval a prompt before you trust it

    Measure an agent or a prompt with real cases and negative controls, then change one thing at a time. Built on foundry-evals and agent-evals-starter.

    Read the steps · 1 min