kit-parity
/kit-parity
Porównuje konfiguracje kilku agentów kodujących na tym samym komputerze i wskazuje różnice w serwerach MCP i umiejętnościach.
- Autor
- fangdididi ↗
- Repozytorium
- kit-parity ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 12.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 5
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 1
Brak sygnałów. Repozytorium zmienione 12.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Porównuje konfiguracje kilku agentów kodujących na tym samym komputerze i wskazuje różnice w serwerach MCP i umiejętnościach.
Wyzwalacze z opisu
Oryginalny opis (EN)
Use when one person runs several AI coding agents (Claude Code, Cursor, Codex, Gemini CLI, VS Code/Copilot, OpenCode, Windsurf, Kiro...) on the same machine or project and wants them to share the same toolkit - the same MCP servers and the same skills - or wants to know why one agent "can do X" and another can't. Inventories every agent's real config read-only, builds a parity matrix, flags drift, inline secrets, shadowed and diverged copies, writes a kit charter, and plans per-agent changes with backups and a load check. Never edits configs without explicit approval.
Jak działa
- What this is
- When to use it
- Ground rules
- Step by step
- Example
- Tips
Podgląd SKILL.md
---
name: kit-parity
description: Use when one person runs several AI coding agents (Claude Code, Cursor, Codex, Gemini CLI, VS Code/Copilot, OpenCode, Windsurf, Kiro...) on the same machine or project and wants them to share the same toolkit - the same MCP servers and the same skills - or wants to know why one agent "can do X" and another can't. Inventories every agent's real config read-only, builds a parity matrix, flags drift, inline secrets, shadowed and diverged copies, writes a kit charter, and plans per-agent changes with backups and a load check. Never edits configs without explicit approval.
---
# kit-parity
**[中文版](SKILL.zh-CN.md)**
## What this is
Most people who use coding agents now use more than one. Each tool keeps its own list of MCP servers and its own skills folder, in its own file format, at its own path. Over a few weeks they drift: a server gets upgraded in one tool and not the others, a skill is edited in one copy, a token gets pasted straight into one config, a project file quietly overrides a user file. The symptom is "it works in Claude Code but Cursor can't see the tracker", and nobody can say which setup is the right one.
kit-parity turns that into a short, checkable job:
1. **Read** every agent's real config without changing anything.
2. **Lay out** a parity matrix: one row per server or skill, one column per agent.
3. **Agree** on a written kit charter: which tools every agent should have, and which are deliberately local.
4. **Plan** the smallest per-agent changes to reach the charter, each with a backup and a rollback.
5. **Prove** each agent actually loads what the charter says, using that agent's own listing command.
The skill is about the toolkit (MCP servers and skills). Instruction files such as `AGENTS.md` or `CLAUDE.md` are only noted, not audited.
## When to use it
- "Make Cursor and Codex have the same MCP servers as Claude Code."
- "Why does this skill behave differently in two agents?"
- You are setting up a new laptop and want the old machine's agent toolkit, minus the junk.
- A team repo ships `.mcp.json` / `.cursor/mcp.json` / `.codex/config.toml` and they disagree.
- Before sharing a dotfiles or config repo, to make sure no token is sitting inline.
Do not use it to install new third-party servers or skills you have not vetted; vet them first, then add them to the charter.
## Ground rules
- **Read-only until approved.** Scanning, matrix and planning change nothing. Writing a config file needs a clear yes for that specific file.
- **Never print a secret.** Report the key name (`Authorization`, `GITHUB_TOKEN`) and where it lives, never the value. Do not copy secrets between configs; move them to environment variables or the tool's own secret store.
- **Respect the tool's own format.** Each agent has its own key (`mcpServers`, `servers`, `mcp`, `[mcp_servers.<name>]`) and its own way of writing environment references (`${VAR}`, `${env:VAR}`, `{env:VAR}`, `*_env_var` fields). Translate, don't paste.
- **Paths change.** Tools move their config files between versions. Treat the built-in path table as a starting point and confirm with the tool's own `mcp list` command or docs.
- **Local can be correct.** A server that exists in one agent only is not automatically drift. The charter decides.
## Step by step
### 1. Frame the job (2 minutes)
Write down in one short block:
- **Agents in scope**: which tools, and on which machine. Include versions if known.
- **Scopes in scope**: user-level, project-level, or both, and which project folder.
- **Reference agent** (optional): the one whose setup the user trusts most.
- **Out of bounds**: files or servers you must not touch (work-managed configs, enterprise policies).
### 2. Inventory, read-only
Run the helper if Python is available:
```bash
python3 scripts/kitparity.py scan --project /path/to/repo
python3 scripts/kitparity.py scan --project /path/to/repo --json > kit-scan.json
```
Without the script, open each config file by hand. For each MCP server record: name, transport (local command or remote URL), command and arguments or URL, the names of environment variables and headers, enabled or disabled, and which file and scope it came from. For each skill folder record: folder name, the `name:` in its frontmatter, whether it is a symlink, and a content hash of the whole folder.
Then ask each agent what it actually loaded (for example `claude mcp list`, `codex mcp list`, the MCP panel in Cursor or VS Code). The file says what is configured; the listing says what is live. Note any difference.
### 3. Read the matrix
The matrix shows one short fingerprint per cell. Same fingerprint means the same definition. Read the findings in this order:
| Finding | Meaning | Usual fix |
|---|---|---|
| `SECRET-INLINE` / `SECRET-IN-ARGS` | A credential value is written in the config | Rotate it if the file was ever shared, then switch to an env reference |
| `UNREADABLE` | The file is malformed, so the tool may be ignoring all of it | Fix syntax first; nothing else in that file counts until then |
| `NOT-ON-PATH` | The launch command is missing for this shell | Install it, or use an absolute path; GUI apps often see a smaller PATH |
| `SERVER-DIFFERS` | Same name, different command, version, args or URL | Pick one in the charter |
| `SHADOWED` | Same server in user and project scope for one agent | Decide which scope owns it; remove the other |
| `SKILL-DIVERGED` | Same skill folder name, different contents | Diff the copies, keep one, symlink or copy it everywhere |
| `NAME-MISMATCH` / `NO-FRONTMATTER-NAME` | Folder name and declared skill name disagree | Some agents load by folder, others by `name:`; make them match |
| `DISABLED` | Present but switched off | Confirm it is intentional |
An absent cell (`.`) is not a finding yet; it becomes one only when the charter says the item belongs there.
### 4. Write the kit charter
The charter is the single statement of what "in parity" means. Keep it as `kit.json` next to the user's dotfiles:
```json
{
"agents": ["claude-code", "cursor", "codex"],
"servers": {
"docs": {"command": "npx", "args": ["-y", "docs-mcp@2.1.0"]},
"tracker": {"url": "https://tracker.example.com/mcp", "only": ["claude-code", "cursor"]}
},
"skills": ["release-notes"]
}
```
Rules for a good charter:
- **Pin versions** of local servers (`pkg@2.1.0`, not `pkg@latest`), so "the same server" really is the same.
- **Use `only`** for items that deliberately live in some agents and not others, and write one line on why.
- **No secret values.** List the environment variable names a server needs in a comment or note instead.
- **One owner scope per server**: say whether it lives at user level or project level.
### 5. Plan the changes
```bash
python3 scripts/kitparity.py check --want kit.json --project /path/to/repo
```
Each line gets a verdict:
- `ADD` - the charter wants it here and it is missing.
- `ALIGN` - it exists but the definition differs from the charter.
- `EXTRA` - present in an agent the charter excludes.
- `UNLISTED` - present but not in the charter at all; ask whether to adopt it or drop it.
- `OK` - matches.
Turn the non-OK lines into a plan, one row per file:
| File | Change | Backup | Rollback |
|---|---|---|---|
| `~/.cursor/mcp.json` | set `docs` args to `docs-mcp@2.1.0` | `mcp.json.bak-2026-11-12` | restore the backup |
Show the exact before and after snippet for each file, written in that tool's own format. Group by file so the user approves each file once.
### 6. Apply (only with approval)
- Copy each file to a dated backup before editing.
- Prefer the tool's own command where it exists (`claude mcp add`, `codex mcp add`) because it writes the format correctly.
- For skills, keep one canonical folder and link or copy it into each agent's skills directory; record which you chose.
- Change one file at a time; re-run `scan` after each.
### 7. Prove it loads
For every agent in scope, open a fresh session and use its own listing or tool panel to confirm each charter item appears and starts. Record a load table:
| Agent | Item | Listed | Started | Note |
|---|---|---|---|---|
A row passes only when the agent itself shows it. "The file looks right" is not proof.
### 8. Hand over
Deliver: the matrix (before and after), the charter, the change log with backup paths, the load table, and a short list of open items (unknown servers, secrets to rotate). Suggest re-running `scan` after any tool upgrade.
## Example
**User:** "Claude Code can create tracker tickets but Cursor says the tool doesn't exist, and Codex has a different docs server. Can you sort this out?"
**Agent:**
1. Frames: three agents, user scope plus the current repo, Claude Code as reference.
2. Scans and finds: `docs` pinned at 2.1.0 in Claude Code and Codex but 1.9.4 in Cursor; the repo's `.mcp.json` also defines `docs` with an extra flag, shadowing the user one in Claude Code; Cursor's `tracker` exists but has a bearer token pasted inline; the `release-notes` skill has different contents in Claude Code and Cursor; Codex has no copy.
3. Says first: "Cursor's tracker config contains an inline token. I won't print it. If this file was ever synced or shared, rotate the token."
4. Drafts the charter with the user: `docs@2.1.0` everywhere at user level, `tracker` only in Claude Code and Cursor, one canonical `release-notes`.
5. Runs `check` and shows the plan: align Cursor's `docs`, move Cursor's token to an env reference, decide whether the project `docs` override stays, add `release-notes` to Codex.
6. After approval, backs up and edits one file at a time, then asks each agent to list its tools, and fills the load table.
## Tips
- GUI-launched agents often do not inherit your shell's PATH or environment variables. A server that works from the terminal can fail in the app; absolute command paths help.
- Remote servers differ in small ways between tools (some need an explicit `"type": "http"`). Compare fingerprints after translating, not the raw JSON.
- Project-level configs are shared with the team; user-level ones are yours. Put personal tokens and experimental servers at user level.
- Some tools only load project configs in trusted folders. If a project item is "missing", check trust before editing.
- Keep the charter in version control without secrets, and the scan output out of it, since it lists local paths.
Pliki w paczce (5)
- LICENSE 1 KB
- README.md 5 KB
- README.zh-CN.md 5 KB
- SKILL.md 10 KB
- SKILL.zh-CN.md 10 KB
Wersje
W bazie nie ma innych skilli o nazwie kit-parity.
Migawka bazy z 13.11.2026 · wersja paczki c33d1a42753a (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100harness-sync
Synchronizuje konfiguracje narzędzi AI w ośmiu celach (Claude Code, Cursor, Copilot, Codex, OpenCode, Windsurf, Gemini CLI, Junie): inwentaryzuje skille, bloki instrukcji i MCP, raportuje rozbieżności i pozwala wypchnąć lub pobrać zmiany. - 100dev-swarm-install-ai-code-agent
Instaluje narzędzia CLI agentów kodujących AI: claude-code, gemini-cli, codex i github copilot-cli. - 100harness-compile
Kompiluje jeden plik harness.yaml do natywnych konfiguracji Claude Code, Cursor, GitHub Copilot, Codex, OpenCode, Windsurf, Gemini CLI i JetBrains Junie, z opcjami --target, --dry-run i --clean. - 95agents-cli
Uruchamia wielu agentów kodujących, takich jak Claude Code, Codex czy Cursor, z jednego CLI, także na kilku kontach. - 80cmk:agent-vendors
Wprowadza jeden zestaw skilli do repozytorium i podłącza go do agentów kodujących, takich jak Claude Code, Codex czy Cursor. - 100ai-context
Generuje, aktualizuje i audytuje pliki kontekstu dla IDE z AI, z kanonicznym AGENTS.md i cienkimi mostkami do Claude Code, Cursor, Copilot, Cline, Windsurf, Gemini CLI, Codex CLI i OpenCode.