Agent Toolkit
Manage reusable skills, personas, workflows, and MCP templates across major coding assistants from one toolkit.
- Source repo
- ulises-jeremias/agent-toolkit
- Stars
- ★ 18
- Last updated
- 3d ago
- License
- MIT
- Primary language
- V
- FA score
- 74/100 · Some gaps
At a glance
- Works with
- Universal · cross-platformChatGPT · Codex · Claude Code
- You'll need
- Typical use
- A development team using Claude Code, Cursor, and Copilot can maintain one capability catalog while deploying each client's native configuration.
- Main limitation
- The broad surface area—skills, personas, loops, plugins, MCP, desktop software, and swarms—creates more setup and governance overhead than a small prompt collection.
- Source review
- 74/100 · Some gaps
What does this agent do, and when should you use it?
Agent Toolkit is a composable capability and runtime layer for AI-assisted software delivery, rather than a standalone model-bound chatbot. Its repository contains skills, agent personas, recurring loops, solution packs, plugins, and MCP templates maintained from shared catalogs and distribution definitions. A native V CLI installs client-specific assets, diagnoses the setup, lists inventory, runs loops, and coordinates multi-agent swarms; a desktop application adds a graphical library, operations view, terminal fleet, and activity insights. Documented targets include Claude Code, Cursor, OpenCode, GitHub Copilot, Windsurf, Pi Agent, and Muse Code, while Agent Plugins 1.0 bundles extend distribution to clients including ChatGPT/Codex. Its outputs stay within the local workspace and target-client configuration boundaries: deployed skill or persona files, native plugin bundles, loop state and reports, and filesystem-backed swarm worktrees and handoffs.
agent-toolkit install detects supported coding assistants and deploys the appropriate skills, personas, and configuration, after which agent-toolkit doctor verifies the setup. agent-toolkit inventory reads the catalogs and capability directories to report the live skill, agent, and loop inventory. Skills are authored as SKILL.md units, personas live under agents/, and product definitions in distributions/products.yaml are compiled into client-facing packages under plugins/. Commands such as agent-toolkit loop run daily-triage execute recurring workflows, write STATE.md and report.md into the loop directory, and apply L1, L2, or L3 mutation-safety policies. agent-toolkit swarm start selects a role recipe and budget, creates a Git worktree for each writer, records state, handoffs, audits, and approvals on the filesystem, and exposes runs through watch and promote. The MCP subsystem supplies registry entries and credential-placeholder JSON templates for GitHub, Slack, Notion, Linear, Figma, ClickUp, and Chrome DevTools.
- A development team using Claude Code, Cursor, and Copilot can maintain one capability catalog while deploying each client's native configuration.
- An open-source maintainer can schedule issue triage, CI sweeps, dependency updates, or PR monitoring with an explicit safety tier for allowed mutations.
- An engineering lead can divide planning, architecture, implementation, review, QA, security, and platform work among narrowly defined personas.
- A team delivering a complex change can use the
pair,team, orfullswarm recipes for isolated worktrees, enforced budgets, audited handoffs, and approval gates. - A developer onboarding a project can start from prepared MCP templates for GitHub, Slack, Notion, Linear, Figma, ClickUp, or Chrome DevTools.
- A user who prefers graphical operations can install the desktop application to browse capabilities, manage jobs and loops, use embedded terminals, and inspect activity reports.
How do you install or deploy this agent?
The recommended product is the native V CLI. With Homebrew:
brew tap ulises-jeremias/homebrew-tap && brew install agent-toolkit
agent-toolkit install
agent-toolkit doctorOn Arch Linux, install the release binary package:
yay -S agent-toolkit-bin
agent-toolkit install
agent-toolkit doctorThe PyPI package is a thin launcher for the bundled V binary:
uv tool install 'agent-toolkit-cli>=1.30.1'
uvx --from 'agent-toolkit-cli>=1.30.1' agent-toolkit install
agent-toolkit doctorThe npm channel is also documented:
npm i -g agent-toolkit-cli
agent-toolkit install
agent-toolkit doctorTo install and launch the desktop application through the CLI:
agent-toolkit gui --install
agent-toolkit guiCore local skills and workflows do not require MCP credentials. If an MCP template is enabled, replace its marked placeholders with credentials for the selected external service.
How do you use this agent?
Install assets for detected clients, verify them, and inspect the available catalog:
agent-toolkit install
agent-toolkit doctor
agent-toolkit inventoryRun the supplied daily triage loop:
agent-toolkit loop run daily-triageInspect swarm recipes and preview a two-role run before allowing execution:
agent-toolkit swarm recipes
agent-toolkit swarm start --recipe pair --dry-run "Add a health check endpoint with tests"Start a team run with the documented Herdr backend, then observe and promote it using the returned run ID:
agent-toolkit swarm start --recipe team --backend herdr "Migrate the auth module"
agent-toolkit swarm watch <run-id>
agent-toolkit swarm promote <run-id>For Agent Skills only, without personas or loops:
npx skills add ulises-jeremias/agent-toolkit -gFor MCP, select a service template from mcp/templates/, replace its credential placeholders, and place the resulting JSON in the target client's MCP configuration location.
What are this agent's strengths and limitations?
- A shared skill and persona catalog targets multiple coding assistants, reducing duplicated prompt configuration across Claude Code, Cursor, Copilot, and other supported tools.
- Loops use explicit L1 proposal-only, L2 controlled-write, and L3 high-autonomy mutation stages, with gates around merge and close operations.
- Swarm orchestration includes role recipes, token/cost/wall-clock budgets, isolated Git worktrees, audited handoffs, bounded feedback rounds, and human approvals.
- Distribution options cover a native CLI, desktop GUI, portable Agent Plugins 1.0 packages, legacy Claude Code plugins, and several system package channels.
- Capabilities, generated products, catalogs, schemas, and validation commands have distinct repository locations, supporting auditable configuration management.
- The broad surface area—skills, personas, loops, plugins, MCP, desktop software, and swarms—creates more setup and governance overhead than a small prompt collection.
- Targets receive different artifacts: Cursor uses plugins or
.mdcrules, Copilot uses an instruction file, and OpenCode uses prompt overlays and agent configurations, so feature parity should not be assumed. profiles/is documented as a deprecated fallback overlay; teams depending on that layout face migration toward the canonicalplugins/output.- MCP templates do not supply service permissions, so adopters must provision and secure credentials for integrations such as GitHub, Slack, Notion, and Linear.
- The full-catalog
agent-toolkit-completepackage is described as experimental and marketplace-pending, which may matter to teams requiring a stable all-in-one distribution.
How does this agent compare with similar options?
Compared with the named classic tmux swarm approach, agent-toolkit swarm is delivered as a single roughly 22 MB V binary and adds built-in pair, team, and full recipes, budget enforcement, a JSON surface, and cross-platform releases. The tmux approach depends on bash, tmux, bb, and a dashboard, although it also provides per-branch packs and first-class handoff auditing. Compared with the repository's legacy Python swarm, the V implementation documents enforced budgets, formal handoff audits, structured observability, and programmatic JSON access without requiring a Python environment.
Key facts side by side with the most closely related agents.
| Agent | Source review | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|
| Agent Toolkit This agent | 74 · Some gaps | ★ 18 | 3d ago | V | ChatGPT · Codex · Claude Code |
| Agentic Plugin Marketplace | 66 · Some gaps | ★ 40k | 3d ago | Python | Codex · Claude Code |
| Agor | 45 · Major gaps | ★ 1.4k | 4d ago | TypeScript | Codex · Claude Code |
| Agency Agents Chinese Expert Library | 28 · Major gaps | ★ 21k | 7d ago | Shell | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The evidence shows mutation tiers, scoped CodeQL permissions, per-user installation, checksum guidance, receipt-oriented uninstall coverage, and environment-variable placeholders for credentials. L3 merge or close operations are described as requiring explicit permission, but the supplied files do not show the gate implementation or per-action confirmation behavior. Telemetry is opt-in and configurable, yet when enabled it exports host data, session identifiers, and truncated raw prompts and responses; this materially limits the sensitive-data score. Dependency controls include commit-pinned Actions, CodeQL, and release verification guidance, but complete lockfiles, audit results, and installer implementations are absent. Attribution, licensing, and the maintainer name are clear; unknown publisher identity was not treated as adverse evidence.
The README, package metadata, security policy, workflow, and launcher tests are broadly consistent about a native V binary with distribution adapters. Missing-binary diagnostics, recovery paths, and propagation of nonzero exit codes are tested concretely, justifying full marks for failure messages. Deductions reflect the absence of executed results, availability claims that remain partly documentary, and the unexplained use of a fixed 1.30.0 acceptance artifact alongside package version 1.32.1.
The project explicitly addresses Claude Code, Cursor, OpenCode, Copilot, Windsurf, Pi, and Muse, with CLI, GUI, plugin, container, offline, and manual installation paths. Boundaries among skills, agents, loops, packs, plugins, and MCP are well described, as are the skills-only npx path and L1/L2/L3 mutation limits. Catalogs, persona responsibilities, and a Jira routing-test harness support trigger design, but one integration's test infrastructure does not establish trigger precision across the full catalog.
Information architecture, directory responsibilities, installation channels, and command examples are extensive. The MIT text agrees with package metadata, and the repository describes semantic releases, a support window, private security reporting, and downstream ownership. Deductions apply because no actual changelog is supplied, FAQ evidence is limited, and limitations are dispersed across platform tests, security policy, and installation notes. Persona archival and several distribution names introduce some naming churn. Response targets are clear, but direct contact depends on an external GitHub profile and no backup maintainer structure is shown.
Install, doctor, inventory, desktop, cross-editor profiles, and staged loops produce outputs that appear directly usable, while compiling one catalog into several tool formats offers plausible marginal value. The strongest benefit claims, including reduced duplication, remain project assertions without user outcomes, comparative evidence, or execution in this review. Cost tracking exists for API-based routing tests, but the supplied material does not provide a user-facing budget model, resource baseline, or cost-benefit comparison across operating modes.
The README ties many claims to directories, catalogs, ADRs, trust documentation, validation scripts, and CI, while package metadata, security policy, tests, and workflows corroborate several distribution and error-handling claims. Many referenced files are not included, so CI status, real-app screenshots, enabled secret scanning, and common provenance of release channels cannot be independently confirmed here. The implementation preserves unknown values in several places, but promotional statements and verified facts are not consistently separated.
- Review the OTLP endpoint and retention policy before enabling test telemetry; the current code can export host identifiers, session IDs, prompts, responses, and code-block content.
- Before granting L2 or L3 loops repository write access, inspect the actual loop-gh-gate implementation, allowlists, confirmation flow, and recovery behavior; the supplied evidence shows the design but not enforcement code.
- Prefer pinned release artifacts with SHA256SUMS, and independently confirm that npm, PyPI, Homebrew, AUR, container, and marketplace adapters resolve to the intended version.
- The npm platform table explicitly covers glibc Linux and excludes musl, so compatibility with Alpine-like environments should not be assumed.
- The security policy states that secret scanning is enabled, but repository settings and historical scan results are not verifiable from this static source set.
FAQ
Does adoption require a particular model provider?
Can a loop modify or merge repository content automatically?
Do MCP integrations require additional permissions?
Can I install only the skills?
npx skills add ulises-jeremias/agent-toolkit -g installs the Agent Skills-compatible content, but it does not include personas or loops.How are broken installations or definitions detected?
agent-toolkit doctor. Contributors can also run ./scripts/validate-skills.vsh and ./scripts/validate-loops.vsh; the validators exit nonzero and print human-readable errors when checks fail.