multi-host-fleet

# Multi-Host Fleet Reference ## What It Does Documents the canonical shell environment variable for every credential-bearing plugin in this marketplace, with patterns for replicating credentials across multiple hosts (workstations, CI runners, ephemeral sandboxes, devcontainers) without re-running `/plugin disable && /plugin enable` per host. The yellow-plugins marketplace adopts a 3-element credential resolution pattern for credential-bearing plugins that bundle a stdio MCP server (yellow-research, yellow-morph, yellow-semgrep): 1. `userConfig` value (stored in OS keychain) — **preferred** for single-host installs 2. Shell env var (canonical name, see table below) — **fallback** for fleet and CI usage 3. Unset → MCP server either fails to start or starts in degraded mode (research, semgrep) Plugins with HTTP-only MCPs (yellow-devin) accept the same userConfig OR shell env var sources, but commands read shell env directly rather than flowing through a wrapper. Setting the canonical shell env var on a new host lets the plugin operate without ever populating the userConfig keychain entry. The enable-time userConfig prompt still fires on platforms where prompting works (see [#39455](https://github.com/anthropics/claude-code/issues/39455)); dismissing or skipping it is safe when the shell env var is set. ## When to Use - Setting up a new workstation, laptop, WSL2 instance, or VM - Configuring a CI job (GitHub Actions, GitLab CI, CircleCI) that runs Claude Code - Building a devcontainer image with pre-injected credentials - Wiring a secrets manager (direnv, 1Password CLI, HashiCorp Vault, Doppler, AWS Secrets Manager) so credentials flow into Claude Code - Diagnosing why a plugin reports PARTIAL or NEEDS SETUP on one host but READY on another ## Usage ### Canonical Env-Var Contract | Plugin | Shell Env Var | userConfig Field | Sensitive | Notes | |--------|---------------|------------------|-----------|-------| | yellow-research | `PERPLEXITY_API_KEY` | `perplexity_api_key` | yes | | | yellow-research | `TAVILY_API_KEY` | `tavily_api_key` | yes | | | yellow-research | `EXA_API_KEY` | `exa_api_key` | yes | | | yellow-morph | `MORPH_API_KEY` | `morph_api_key` | yes | | | yellow-semgrep | `SEMGREP_APP_TOKEN` | `semgrep_app_token` | yes | `sgp_` prefix | | yellow-devin | `DEVIN_SERVICE_USER_TOKEN` | `devin_service_user_token` | yes | HTTP MCP; commands read shell env directly | | yellow-devin | `DEVIN_ORG_ID` | `devin_org_id` | no | HTTP MCP; commands read shell env directly | The userConfig path uses the system keychain when available (macOS, Windows) or `~/.claude/.credentials.json` (0600 perms) on minimal Linux. The shell env path is what makes multi-host fleets practical. ### Dev hosts (dotfiles + shell rc + direnv) For a typical developer workstation, store secrets in shell rc files or direnv: ```bash # In ~/.zshrc or ~/.bashrc (NOT version-controlled in plaintext) export PERPLEXITY_API_KEY="$(cat ~/.secrets/perplexity 2>/dev/null)" export TAVILY_API_KEY="$(cat ~/.secrets/tavily 2>/dev/null)" export EXA_API_KEY="$(cat ~/.secrets/exa 2>/dev/null)" export MORPH_API_KEY="$(cat ~/.secrets/morph 2>/dev/null)" export SEMGREP_APP_TOKEN="$(cat ~/.secrets/semgrep 2>/dev/null)" export DEVIN_SERVICE_USER_TOKEN="$(cat ~/.secrets/devin-token 2>/dev/null)" export DEVIN_ORG_ID="your-org-id" ``` Or with [direnv](https://direnv.net) (per-project `.envrc`, not committed): ```bash # .envrc (gitignored) export PERPLEXITY_API_KEY="$(cat ~/.secrets/perplexity 2>/dev/null)" # ... etc ``` `~/.secrets/` files at 0600 perms are the simplest portable secret store. For shared workstations or higher-sensitivity environments, use the secrets-manager patterns below. ### CI/CD (GitHub Actions, GitLab CI, devcontainer) **GitHub Actions:** ```yaml - name: Run /flow:work env: PERPLEXITY_API_KEY: ${{ secrets.PERPLEXITY_API_KEY }} EXA_API_KEY: ${{ secrets.EXA_API_KEY }} SEMGREP_APP_TOKEN: ${{ secrets.SEMGREP_APP_TOKEN }} run: claude -p "/flow:work ..." ``` Add each canonical env var name to **Repository Settings → Secrets and variables → Actions**. Sensitive values go in `secrets`. **GitLab CI:** Define each canonical env var as a Project CI/CD variable (mask sensitive ones). They auto-propagate to job environments. **devcontainer.json:** ```json { "containerEnv": { "PERPLEXITY_API_KEY": "${localEnv:PERPLEXITY_API_KEY}", "EXA_API_KEY": "${localEnv:EXA_API_KEY}", "SEMGREP_APP_TOKEN": "${localEnv:SEMGREP_APP_TOKEN}" } } ``` Pulls from the host's env (which must be set per the dev hosts section above). For ephemeral cloud devcontainers (Codespaces, Coder), inject via the platform's secret-management UI. ### Secrets managers (tool-agnostic) **direnv (recommended for local dev — simplest):** ```bash # .envrc (per-project, gitignored) dotenv ~/.secrets/yellow-plugins.env # ~/.secrets/yellow-plugins.env (0600 perms, gitignored) PERPLEXITY_API_KEY=pk_... EXA_API_KEY=... ``` direnv loads env vars when you `cd` into the project. Use `direnv allow` to authorize the `.envrc`. **1Password CLI (`op run`):** ```bash # .envrc.template (committed) export PERPLEXITY_API_KEY="op://Personal/Perplexity API/credential" export EXA_API_KEY="op://Personal/EXA API/credential" # Then launch Claude Code via: op run --env-file=.envrc.template -- claude ``` **HashiCorp Vault (envconsul):** ```bash envconsul -vault-addr=https://vault.internal -secret=secret/yellow-plugins claude ``` **Doppler:** ```bash doppler run --project yellow-plugins -- claude ``` **Generic env-file pattern (any tool):** The contract is: by the time `claude` starts, the canonical env vars must be in the process environment. Any tool that achieves this works. The wrapper scripts inside each plugin will pick them up. ## Examples ### Complete `.zshrc` snippet (uncomment what you use) ```bash # === yellow-plugins env-var contract === # Source of truth: plugins/yellow-core/skills/multi-host-fleet/SKILL.md # yellow-research # export PERPLEXITY_API_KEY="$(cat ~/.secrets/perplexity 2>/dev/null)" # export TAVILY_API_KEY="$(cat ~/.secrets/tavily 2>/dev/null)" # export EXA_API_KEY="$(cat ~/.secrets/exa 2>/dev/null)" # yellow-morph # export MORPH_API_KEY="$(cat ~/.secrets/morph 2>/dev/null)" # yellow-semgrep # export SEMGREP_APP_TOKEN="$(cat ~/.secrets/semgrep 2>/dev/null)" # yellow-composio and yellow-linear use browser OAuth. They have no # shell env credential. Headless Composio uses `claude mcp add` with a # For You consumer key; see /composio:setup. # yellow-devin # export DEVIN_SERVICE_USER_TOKEN="$(cat ~/.secrets/devin 2>/dev/null)" # export DEVIN_ORG_ID="your-org-id" ``` ### Verifying credentials resolve correctly After setting env vars, run `/setup:all` to verify each plugin reports READY. Look at the "Credential Status Files" section — each credential-bearing plugin should show `source: shell_env` (not `absent`) for fields you set via env. If a plugin shows the OLD source (e.g., `source: userConfig` from a previous keychain entry) but you want the shell env value to win, remember: **userConfig wins over shell env**. To prefer shell env, run `/plugin disable <name>` then `/plugin enable <name>` and dismiss the userConfig prompt (or unset the keychain entry). ## Security - Never commit `.env`, `.envrc`, or `~/.secrets/*` files. Use `.gitignore` patterns: `.env`, `.envrc`, `*.secret`, `!.envrc.template`. Do not use a bare `.env*` glob — it also matches `.envrc.template`, which this guide recommends committing. - Keep `~/.secrets/` at 0700 directory perms and 0600 file perms. - For sensitive plugins (semgrep, perplexity, exa, tavily, morph), prefer keychain (userConfig) over shell env when running on a single host. Shell env is the only practical option for fleets. - yellow-composio authenticates with Claude Code's browser OAuth against `https://connect.composio.dev/mcp`. It is not in the env-var table. - Env vars are visible to MCP subprocesses spawned by Claude Code. Don't store keys with broader scope than needed.