Esper

Esper is a Claude Code and Codex plugin marketplace. Seven development plugins published through synchronized Claude Code and Codex catalogs: idea-to-spec elicitation interviews, multi-agent code review, queryable architecture maps, parallel plan execution, spec consolidation, vulnerability-aware dependency upgrades, and a generated codebase wiki.

The repository carries two synchronized catalogs:

• .claude-plugin/marketplace.json for Claude Code
• .agents/plugins/marketplace.json for Codex

Each plugin is implemented in its own repository. A catalog entry becomes installable in Codex once that source repository ships a valid .codex-plugin/plugin.json; until then, Codex may omit or reject that individual entry while the Claude package remains available.

Installation

Claude Code

# One-time: add the marketplace
claude plugin marketplace add MisterVitoPro/esper

# Then install whichever plugins you want (see per-plugin sections below)
claude plugin install qa-swarm@esper

---

Codex

# One-time: add the marketplace
codex plugin marketplace add MisterVitoPro/esper

# Inspect available entries, then install a Codex-enabled plugin
codex plugin list
codex plugin add <plugin-name>@esper

Start a new Codex session after installing so bundled skills and tools are loaded. If a plugin includes lifecycle hooks, review and trust them with /hooks before expecting them to run.

---

Plugins

qa-swarm

A swarm of specialist agents reviews your code, then fixes the bugs test-first.

Claude Code and Codex ready. Version 1.5.1 ships both manifests, Codex-valid skills, host-neutral user gates, and portable bundled-role loading for native subagent dispatch.

Six Sonnet core reviewers (security, correctness, performance, architecture, data-flow, async) run in parallel, optionally joined by up to six Haiku specialists (config/env, supply chain, type safety, state mgmt, logging, backwards-compat). Findings are deduplicated, ranked P0–P3, and corroborated across agents. The implement skill picks up the report and fixes issues with a 3-agent TDD loop (failing test → minimal fix → verify).

# Claude Code
claude plugin install qa-swarm@esper
/qa-swarm:attack "audit the authentication flow for security and correctness"
/qa-swarm:implement

# Codex
codex plugin add qa-swarm@esper
$qa-swarm:attack "audit the authentication flow for security and correctness"
$qa-swarm:implement

→ Plugin docs

---

code-atlas

Stops Claude Code and Codex from re-exploring your repo every session.

Claude Code and Codex ready. Version 2.3.2 reloads the index after /clear and compaction instead of only at startup, and dispatches its bundled analysts through native agent types on Claude Code while embedding the same definitions on Codex. Codex users must review and trust the bundled hook with /hooks before automatic index injection begins.

Three analyst agents scan your codebase in parallel and a graph synthesizer annotates the key modules, producing a curated architecture index (directory map, key files, tech stack, patterns, dependencies, build commands) plus a semantic dependency graph. Graph queries (dependencies, dependents, blast radius, risk filters) run deterministically through a bundled Node script. A SessionStart hook injects the index as context on startup, resume, /clear, compaction, and forks, and both the map and update skills upsert a short marker-delimited usage block into CLAUDE.md / AGENTS.md so later sessions consult the atlas instead of grepping from scratch. Incremental updates re-scan only what changed.

# Claude Code
claude plugin install code-atlas@esper
/code-atlas:map                   # full first-time scan
/code-atlas:update                # incremental refresh
/code-atlas:query "what calls AuthService.login?"

# Codex
codex plugin add code-atlas@esper
$code-atlas:map
$code-atlas:update
$code-atlas:query "what calls AuthService.login?"

→ Plugin docs

---

plan-runner

Take a Markdown implementation plan, run it through a parallel agent swarm.

Claude Code and Codex ready. Version 1.16.1 pipelines verification off the critical path and loads bundled analyzer, developer, verifier, and aggregator roles relative to the active skill. Codex uses native subagents; Claude Code can additionally opt into Agent Teams.

The analyzer breaks your plan into file-disjoint waves (≤6 agents per wave). Dev agents implement tasks in parallel; each wave commits, then its verifier checks acceptance criteria against a snapshot of that commit while the next wave's agents are already running -- every verdict still lands before aggregation, and --sync-verify restores blocking verification. The aggregator dedupes bugs and emits a fix-plan.md for re-runs. Per-wave git commits keep history bisectable. TDD red-green mode is on by default (--no-tdd to disable), with one shared full-suite run per wave for the regression diff. Large plans (over 4 waves) split into sequential phases checkpointed to run-state.json; --resume picks an interrupted run back up at the last completed wave. Claude Code can auto-detect Agent Teams (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1, v2.1.178+); other Claude sessions and Codex use native subagents. Every dispatched subagent's token usage is tallied best-effort into manifest.json, with explicit lower-bound coverage when the host exposes incomplete usage data. At the end it can push the branch and open or update a structured pull request through the bundled PR skill.

# Claude Code
claude plugin install plan-runner@esper
/plan-runner:run path/to/implementation-plan.md

# Codex
codex plugin add plan-runner@esper
$plan-runner:run path/to/implementation-plan.md

Pairs with the ideas plugin as the pipeline's front door: its interview skill produces an audited spec and emits a plan-runner-ready plan for this skill to execute.

→ Plugin docs

---

jupiter

Consolidates scattered specs into a single canonical tree, and stubs out the docs you forgot to write.

adopt reorganizes spec files in place, grouped by module (multi-module repos) or feature (single-module). rewrite collapses everything to one file with optional source cleanup. The surface scanner walks your code and appends stubs for undocumented agents, skills, CLIs, and configs so nothing slips through.

claude plugin install jupiter@esper
/jupiter:adopt          # reorganize specs in place
/jupiter:rewrite        # collapse to single master file

→ Plugin docs

---

migration-runner

Vulnerability-aware dependency upgrade orchestrator across 7 ecosystems.

Scans for outdated packages (npm, Python, Go, Rust, Java, Kotlin, C#), queries OSV.dev for CVEs, recommends the safest-yet-most-recent target version per package, then executes wave-by-wave with build/typecheck/test verification and clean git rollback on failure. The two-step flow (detect then run) lets you review the plan before any code is touched.

claude plugin install migration-runner@esper
/migration-runner:detect
/migration-runner:run

→ Plugin docs

---

llm-wiki

Turns your codebase into a navigable wiki -- written for both new engineers and Claude's per-task context.

A planner decides the page set, writer agents fill one page each in parallel waves, a diagram author derives Mermaid diagrams from the dependency graph, and a synthesizer builds a session-loaded index with validated cross-links. It is the prose layer that complements code-atlas: where code-atlas is a machine-first dependency graph, llm-wiki writes the human-and-agent-readable "how and why" -- consuming the code-atlas index as ground truth when present, else self-scanning. Pure static Markdown (no embeddings), per-page source_files provenance, and git-blob hash-diff staleness detection that regenerates only stale pages. A SessionStart hook loads the index so Claude reads one page per task instead of grepping.

claude plugin install llm-wiki@esper
/code-atlas:map                   # optional but recommended -- llm-wiki reuses the graph
/llm-wiki:generate                # build the wiki under .llm-wiki/
/llm-wiki:update                  # incrementally refresh stale pages

→ Plugin docs

---

ideas

Interviews you into an audited design spec before any code gets written — then hands plan-runner its input.

Claude Code and Codex ready. Version 0.7.2 ships both client manifests, Codex-valid skills, host-neutral structured-question handling, and portable auditor/critic subagent loading. Choosing "Execute with plan-runner" at the plan completion gate now hands off instead of invoking in-session: it tells you to /clear first, then prints the exact plan-runner:run command to paste, so the execution engine starts with a fresh context and only the plan file.

A scope-sized interview (S/M/L triage, batched multiple-choice waves, hard cap of 5 question calls) pins the existing-system baseline from your repo first, then records every answer in an on-disk ledger — decided / assumed / open — so the run survives /clear and resumes from the file alone. A category-coverage elicitation floor sweeps the ambiguity taxonomy (non-functionals, lifecycle, and interfaces weighted first) so interviews can't close with critical ground unasked, and every unconfirmed item becomes a binding default welded into an acceptance criterion or a blocking open question — never a passive flag a builder can ignore. Two read-only agents gate the draft: a binding ledger audit (every spec claim traces to a decision; a model guess is never recorded as a user decision) and a biggest-miss critic. Output: a committed spec with EARS acceptance criteria, brownfield change deltas, and optional MADR-lite ADRs that later interviews read to skip already-decided questions. After approval, "Approve + generate plan" (or /ideas:plan run standalone against an approved spec) emits a plan-runner-ready plan — contracts only, full criterion text per task — and /ideas:tickets projects it to GitHub as a parent tracking issue plus one linked sub-issue per task behind a Definition-of-Ready gate (gh CLI only). It complements plan-runner, it does not replace it: interview -> spec -> plan here, execution there. Interview behavior is benchmark-tuned against a paired simulated-user harness (ideas-bench).

# Claude Code
claude plugin install ideas@esper
/ideas:interview "your rough idea here"
/ideas:plan docs/specs/approved-spec.md
/ideas:tickets docs/plans/approved-spec.plan.md

# Codex
codex plugin add ideas@esper
$ideas:interview "your rough idea here"
$ideas:plan docs/specs/approved-spec.md
$ideas:tickets docs/plans/approved-spec.plan.md

→ Plugin docs

---

Troubleshooting

For Claude Code, if /plugin shows errors after installing:

claude --debug                      # shows plugin load errors
/plugin                             # opens the plugin manager UI

code-atlas, plan-runner, and llm-wiki ship Node.js SessionStart hooks; Node must be on PATH.

For Codex, inspect the configured marketplace and plugin state:

codex plugin marketplace list
codex plugin list

If a catalog entry is missing, confirm that its pinned source contains .codex-plugin/plugin.json. Start a new session after installing or updating a plugin.

Contributing

1. Create a standalone plugin repository with both .claude-plugin/plugin.json and .codex-plugin/plugin.json, plus its agents, skills, hooks, scripts, and documentation as applicable.
2. Register the same URL, release ref, and commit sha in .claude-plugin/marketplace.json and .agents/plugins/marketplace.json. Add the required Codex installation/authentication policies and category.
3. Tag the release as plain v<version> in the plugin repository, then update both catalog pins together. CI rejects catalog drift.
4. Keep client-specific surfaces honest: Claude-only features may remain documented as such until an equivalent Codex skill, MCP server, or hook is provided.

See CLAUDE.md for repo conventions.

License

MIT. See LICENSE.