Summon is a drop-in framework for Claude Code. It installs agent definitions, slash commands, and process docs into your project. No runtime. No dependencies. Just markdown.
A lone agent says “looks clean, shipping.” Summon is the engineering team that asks “who decided this?” and holds the merge until someone has.
Before — a capable agent, alone
A Claude session set out to improve Summon, and did good work: four clean commits, tests passing, diffs that read well. It shipped. What no one wrote down was that four of those commits had rewritten how the team operates — its review gates, its definition of done — with no decision record behind them. No ADR. No second opinion.
After — run through the team
Seven agents reviewed it, each with standing to halt. Wei named the omission: the team had rewritten its own governance with no record. Archie reclassified the four commits from “done” to a spike to ratify or revert, and the work stayed frozen until two ADRs were written, argued across three rounds, and ratified by a human.
The uncomfortable part is the actual point: this happened in the Summon repo, by an agent that had every gate available and skipped them anyway. Gates you have to remember to invoke get skipped; that’s human, and prompt-tuning a single agent doesn’t fix it, because nothing in a solo pass is built to notice the absence of a decision. What Summon changes is the aftermath: the omission became legible and reversible, and ADR-0004 and ADR-0005 now exist. Judge that, not a promise that nothing ever slips.
Read the receipt: the full episode — the verdicts, the freeze, the two ratified ADRs — is recorded in the review log.
Architecture gates
No implementation without an ADR. The calls that are expensive to reverse get a decision record before the code, not after.
Vetoes with teeth
Pierrot can halt a merge on security. Tara can halt it on test coverage. The team says no when it should, and the no actually stops the work.
Multi-lens review
Simplicity, test quality, security, and architecture, reviewed in parallel and scored independently — never blended into one agreeable summary.
A decision trail
ADRs written before the code they govern. Six months on, “why is it like this, and who agreed?” has an answer in the git log.
A newly added dependency has to age at least three days before Summon adopts it. Most supply-chain compromises are caught and yanked within days, so you’re usually not the one who installs the poisoned version.
Release-age cooldown
New or upgraded dependencies must be at least three days old before adoption, unless a human logs the override. Prefers the package manager’s native minimumReleaseAge, which ages transitive dependencies too where available.
SHA-pinned CI
CI actions pinned to exact commit SHAs, not mutable tags. A v4 tag can be silently repointed to code you never reviewed. A lockfile-drift guard flags dependency files that shift with no commit behind them.
MCP-server vetting
Review a server before adding it, pin its version, keep an allowlist, and watch for rug-pulls, where a trusted server ships a hostile update.
Owner-harm threat lens
The agent holds your shell. STRIDE models harm to your users; the C1–C8 lens models harm to you, its operator: rm -rf on real data, a leaked credential, an action nobody authorized.
You’re the product owner and the final decision-maker. The team handles the engineering discipline.
npx summon-team my-projectcd my-project# Open in Claude Code, then run /quickstartThat scaffolds the framework into a new project. Open it in Claude Code and run /quickstart to begin.
Summon installs 16 agents and 24 slash commands as plain markdown. The full roster and the command reference live in the docs:
/kickoff to /handoff