Skills & MCP Enhancements
KAIROS agents declare optional enhancements via skills and MCP tools. The framework works fully without them — all enhancements are additive only. No skill or MCP is a hard dependency.
KAIROS does not recommend or depend on any third-party Skill marketplace. Every enhancement below is either an Anthropic-authored Claude Code built-in, an internal skill authored and shipped as part of the KAIROS plugin itself (auto-installed with claude plugin install kairos, same as every agent file), or an MCP server.
How a Skill Actually Works
A Skill is one Markdown file (SKILL.md) with a YAML description header and an instruction body — a checklist, a format, a process. It is not injected into every prompt: Claude Code's Skill tool loads it on demand, only when the current task matches that description, and only into the agent that invoked it. This is different from an MCP server, which is a live external tool an agent calls repeatedly (a browser, a database, a deploy target) — a Skill is read-once instructions, an MCP is a running connection.
KAIROS ships its own internal skills, under skills/ in this repo, for one reason: several agents needed the identical checklist (how to tally a Risks table, what error format to show on missing input) and duplicating that prose across 9+ agent files meant it would drift out of sync the first time one file got edited and the others didn't. Each internal skill lives once; every consuming agent's body just says "apply the <name> skill" and the shared text loads at that point instead of being copy-pasted per agent.
MCP Tools
| MCP | Stars | Install | Agents that benefit |
|---|---|---|---|
| Chrome DevTools MCP | 42.8k | See install guide | test-verifier-agent, teammate-tests-agent, teammate-frontend-agent |
| Playwright MCP | 33.3k | See install guide | test-verifier-agent, teammate-tests-agent, teammate-frontend-agent |
Chrome DevTools MCP enables screenshot capture, console inspection, network monitoring, Lighthouse audits, and custom in-browser assertions. Playwright MCP enables full E2E test execution against a running app.
Subagent tool access requires an explicit frontmatter grant
Installing the MCP server is not enough. test-verifier-agent, teammate-tests-agent, and teammate-frontend-agent run as Claude Code subagents (via the Agent/Task tool), whose tools: frontmatter is a strict allowlist — a tool not listed there is unreachable even if the MCP is connected session-wide, and even if the agent's body prose tells it to use that tool. agents/test-verifier-agent.md, agents/team/teammate-tests-agent.md, and agents/team/teammate-frontend-agent.md all list mcp__chrome-devtools__*, mcp__playwright__* in tools: for exactly this reason — if you add MCP tools to another agent, grant them the same way (mcp__<server>__* wildcards a whole server) or the agent will silently never call them.
Skills
Built-in Claude Code skills
No install required — these ship with Claude Code.
| Skill | What it does | Agents that benefit |
|---|---|---|
code-review | Structured multi-angle review pass (correctness, security, style, standards) | code-reviewer-agent |
security-review | Adversarial security audit checklist — asks "how do I break this," not just "does this pass a compliance box" | security-reviewer-agent, teammate-backend-agent |
verify / run | Actually executes the project (build, test, run) to confirm a change works, instead of trusting a static read of the code | implementer-tdd-agent, implementer-coder-agent, test-verifier-agent, teammate-tests-agent, teammate-frontend-agent, release-planner-agent |
deep-research | Multi-source investigation pattern for open-ended questions, before committing to an answer | context-extractor-agent, impact-assessment-agent, pm-agent, architect-agent, retrospective-agent |
User-installed skills
Not shipped with Claude Code and not part of the KAIROS plugin — a personal or third-party skill the user installs themselves. KAIROS never requires it.
| Skill | What it does | Agents that benefit |
|---|---|---|
issues-generator | Drafts GitHub/GitLab issue text from a requirement, in either an implementation-ready or an outcome-driven mode | pm-agent |
Internal (KAIROS-authored) skills
No install required — these ship as part of the KAIROS plugin itself, under skills/ in this repo, and are auto-installed together with the agent files by claude plugin install kairos. Each row links to the actual SKILL.md — open it to see the full checklist, not just the summary below.
| Skill | What it enforces | Agents that use it |
|---|---|---|
agent-contract | One shared 🚨 AGENT ERROR format for a missing required input, instead of nine agents each inventing their own wording | pm-agent, architect-agent, impact-assessment-agent, code-reviewer-agent, security-reviewer-agent, test-verifier-agent, qa-plan-agent, release-planner-agent, documentation-agent, retrospective-agent |
analysis-discipline | Review-side counterpart of coding-discipline: evidence-backed findings, restraint on low-value nitpicks, scope-bounded investigation, brief direct pushback when the evidence contradicts the requested approach, and a sweep of the same defect class across the enclosing unit before any finding is recorded, pre-existing occurrences included | context-extractor-agent, impact-assessment-agent, bug-triage-agent, pm-agent, architect-agent, code-reviewer-agent, security-reviewer-agent, test-verifier-agent, qa-plan-agent, dependency-audit-agent |
constraint-taxonomy | The closed Category vocabulary of ledger/constraints.md, written once at row creation and never rewritten, plus the reader/writer/gating rules that let a review section run only when its obligation was declared upstream — an obligation is never inferred from the code | every agent that creates or updates a constraint row, plus code-reviewer-agent (Accessibility), security-reviewer-agent (Compliance & Privacy) and qa-plan-agent (Cross-Environment Verification), which are gated on it |
contract-checklist | 9 questions to resolve before finalizing any API/DB contract — entity lifecycle, IDOR risk, idempotency, delete behavior, pagination, error shape | architect-agent, implementer-lead-agent |
threat-model | Design-time STRIDE-lite: trust boundaries, attack surface added, who holds authority, what the attacker controls, data crossing the boundary, failure and abuse, blast radius — findings land in the architecture artifact's own Risks table | architect-agent (Step 5b) |
migration-safety | Expand/contract sequencing, lock duration at production row counts, idempotent and restartable backfill, reversibility, what a code rollback does to data already written in the new shape, ordering against the code deploy | architect-agent (when the data model changes), release-planner-agent (rollback strategy) |
code-simplification | REFACTOR-step checklist: what to simplify, how to confirm behavior didn't change, when to stop | implementer-tdd-agent, implementer-coder-agent (REFACTOR step) |
test-scope | How much of the suite to run at each stop: test files, related tests or the whole suite, which stage owns the full run, how a targeted run reports coverage, the test-changed setting, and the rules for writing tests that are cheap to run | implementer-tdd-agent, implementer-coder-agent, test-verifier-agent, implementer-lead-agent |
artifact-bookkeeping | Pure-arithmetic rules for tallying a Risk/Issue/Finding table by impact and deriving each phase's pass/fail status from a fixed threshold — no agent "eyeballs" a count | pm-agent, architect-agent, impact-assessment-agent, both implementers, code-reviewer-agent, security-reviewer-agent, test-verifier-agent, qa-plan-agent, release-planner-agent, documentation-agent, orchestrator-agent |
artifact-template | The mandatory 4-line ## Summary head block every phase artifact opens with, and the fixed column sets for any table the Risk Disposition Loop parses — so a gate is readable in seconds and no two phases spell Mitigation/Fix differently | every agent that writes a .kairos/ phase artifact, plus orchestrator-agent (reads the Summary at each gate) |
coding-discipline | Pre-implementation checklist: scope discipline, no speculative abstraction, surface assumptions, trust boundaries, WHY-only comments | implementer-tdd-agent, implementer-coder-agent, implementer-lead-agent, all 4 layer agents (teammate-*-agent) |
issue-tracker-comment | Shared jira/glab/bitbucket commands for posting a phase's output as a comment on the originating issue — whole file, title-prefixed, or the extract body qa-plan-agent sends a tester (in an orchestrated run the orchestrator posts it, after the QA gate) | every phase agent's own optional tracker-comment step |
These replace what used to be third-party plugin dependencies (karpathy-guidelines and 11 Trail of Bits plugins) — see CHANGELOG.md for the removal. Their guidance was folded into each consuming agent's own checklist natively (a concrete check item, an unconditional process step) rather than gated behind "if this external skill is installed."
Full agent × enhancement map
| Agent | Phase | Skills | MCP |
|---|---|---|---|
orchestrator-agent | Coordinator | — | — |
context-extractor-agent | Pre | deep-research | — |
impact-assessment-agent | Pre | deep-research | — |
bug-triage-agent | Pre (bugs) | deep-research | — |
pm-agent | 1 | deep-research, issues-generator | — |
architect-agent | 2 | deep-research | — |
implementer-tdd-agent | 3a | verify/run | — |
implementer-coder-agent | 3b | verify/run | — |
code-reviewer-agent | 4 | code-review, security-review | — |
security-reviewer-agent | 4 | security-review | — |
test-verifier-agent | 5 | verify/run | Chrome DevTools MCP, Playwright MCP |
qa-plan-agent | 5b | — | Chrome DevTools MCP (optional) |
release-planner-agent | 6 | verify/run | — |
documentation-agent | 6b | — | — |
retrospective-agent | Post | deep-research | — |
improvement-advisor-agent | Post | — | — |
dependency-audit-agent | Standalone | deep-research | — |
implementer-lead-agent | Layered | — | — |
teammate-backend-agent | Layered | security-review | — |
teammate-database-agent | Layered | — | — |
teammate-frontend-agent | Layered | verify/run | Chrome DevTools MCP, Playwright MCP |
teammate-tests-agent | Layered | verify/run | Chrome DevTools MCP, Playwright MCP |
Every agent above also applies whichever internal (KAIROS-authored) skill from the table above fits its role — omitted here since this table tracks optional enhancements, and the internal skills are a native part of each agent's own process, not a bolt-on. agent-contract specifically is omitted from every row above for the same reason and because it applies almost universally (see the internal-skills table above for its full agent list), not per-phase.
Gaps (no generic solution available)
- Database MCP — no mature generic database MCP exists. All database MCPs are vendor-specific.
teammate-database-agenthas no MCP enhancement. - Deploy MCP — all deploy MCPs (Vercel, Buildkite, etc.) are vendor-specific.
release-planner-agentusesverify/runfor smoke testing but has no generic deploy automation. - LSP — useful for type-aware code navigation, but no implementation with sufficient adoption exists (all available LSP MCP wrappers have under 400 stars). Excluded pending a more mature option.