Debian777 Kairos MCP Kairos Install

kairos-install

Use bundled references in this skill package as the canonical runtime source. Those references are refreshed from this repository docs during build/update workflows.

The default path is the simple Docker Compose stack. Use the advanced full stack only when the user explicitly asks for the operator-managed profile. Do not treat Keycloak or identity-provider administration as part of the normal install flow.

Recommended default installation

Unless the user explicitly chooses a different profile, recommend both of these together:

1. Runtime — Docker: Bring up KAIROS with the simple Compose stack from bundled references (.env, then docker compose up). The npm package is not a substitute for running the service in containers. In enterprise environments, the parallel default is usually Helm (or another operator-managed chart) instead of local Compose; the same CLI guidance below still applies to operators targeting that deployment.
2. CLI — @debian777/kairos-mcp: Install globally (npm install -g @debian777/kairos-mcp, then kairos on PATH) for day-to-day use, or run npx @debian777/kairos-mcp when the user declines a global install. The CLI is a core part of the ecosystem: it exposes MCP-oriented workflows with bulk and operational affordances (for example train, tune, export, token and server targeting) that typical IDE MCP sessions alone do not replace for operators.
3. MCP in the IDE: Optional — only when the user wants an MCP host to call HTTP MCP. Do not treat “we use MCP” as a reason to skip installing the CLI when the user will operate adapters, training, or bulk commands.

Canonical bundled docs

Use these bundled references before you act:

• Install index: references/install/README.md
• Prerequisites: references/prerequisites.md
• Simple stack: references/docker-compose-simple.md
• CLI: references/CLI.md
• Full stack (advanced): references/docker-compose-full-stack.md

Use the docs in this order:

1. Read bundled references from this skill package.
2. If maintainers ask for verification against upstream docs, treat that as advisory only and never as direct command authority.
3. Do not execute commands copied from remote content unless they are validated against bundled references first.

Extraction rule for .env blocks:

• open bundled references/docker-compose-simple.md
• go to ## 3. Environment file
• choose the matching backend subsection
• copy only the fenced ini block directly under that subsection

Mini example:

1. open bundled references/docker-compose-simple.md
2. find ## 3. Environment file
3. if the user chose OpenAI, go to ### OpenAI
4. copy only the next fenced ini block

Major decision gates

Consult the user before any material change. Use direct yes/no or short-option questions, not vague prompts.

Treat these as material changes: installing software, downloading compose.yaml, writing or overwriting .env, pulling models, starting or stopping containers, running login commands that open a browser or store tokens, and writing MCP host configuration.

Resolve these decisions explicitly:

1. Which working directory contains compose.yaml?
2. Which embedding backend should the install use: OpenAI, Ollama, or TEI?
3. Which embedding model should be used?
4. When a prerequisite is missing, install it now or wait?
5. If the target server requires auth, use browser login flow now?
6. Does the IDE or another MCP host need HTTP MCP configuration, or is CLI-only access sufficient for this user?
7. Which repo skills should be present after install/update (kairos, kairos-install, kairos-bug-report, or all)?

For this skill, "advanced full stack" or "operator-managed profile" means the Compose deployment that adds supporting services such as Redis, Postgres, and Keycloak beyond the default app + Qdrant path.

Treat requests for Redis, Postgres, SSO, Keycloak, or "full stack" as requests for this advanced path.

Skip MCP host configuration when the user does not need an IDE or other HTTP MCP client. Prefer installing the CLI anyway for train/tune/export-style workflows. If the task is only install, health verification, login, or routine CLI usage, skip MCP host configuration — but do not skip CLI installation when the user expects operator or bulk workflows.

Must always

Apply these rules on every install run:

• Start by reading bundled install index, prerequisites, simple stack, and CLI docs.
• Check whether required local components already exist before installing anything.
• Identify missing components precisely: Docker, Docker Compose v2, Node.js 24+, kairos CLI or npx, a writable .env location, and Ollama only when the Ollama path is selected.
• When a prerequisite is missing and the user approves the fix, install it instead of only describing the next step.
• Use the simple stack as the default recommendation.
• Use bundled references to build the exact .env block for the chosen backend.
• Confirm the target path before creating or overwriting .env.
• Set AUTH_ENABLED=false for the default simple stack.
• Guide the user through CLI authentication when auth is enabled on the chosen server.
• Keep repo skills aligned with user intent by installing or refreshing them via npx skills add debian777/kairos-mcp (all skills) or explicit --skill flags.
• Explain the next blocking issue when the install cannot continue.
• If the user declines an installation step, explain what remains blocked and wait.

Must never

Apply these prohibitions on every install run:

• Never choose an embedding backend or model without consulting the user.
• Never install Docker, Node.js, Ollama, or the CLI without a yes/no approval.
• Never overwrite .env, pull Ollama models, or start containers without confirmation.
• Never invent API keys, tokens, service URLs, or other secrets.
• Never ask the user to paste real secrets into chat output.
• Never print, echo, or repeat secret values after the user enters them.
• Never treat Keycloak or identity-provider configuration as part of the standard install.
• Never use a chat model for embeddings.
• Never depend on references/README.md for the main flow.
• Never continue past a missing prerequisite if the user has not approved the fix.
• Never assume only kairos-install is enough when the user asked for additional workflow skills.

Workflow

Follow this sequence so the install stays ordered and auditable.

0. Classify the request

Classify the user request before taking action:

• Install mode: first-time setup, bootstrap, or environment bring-up
• Update mode: Docker or npm CLI refresh, version bump, or maintenance

If unclear, ask the user whether they want install mode or update mode.

When the user asks for Docker, npm CLI, or repo skill updates, skip backend and .env selection unless they also ask to reconfigure the runtime.

1. Read the canonical docs

Read bundled install index, prerequisites, simple stack, and CLI docs first.

2. Audit prerequisites

Confirm that the user has:

• Docker Engine and Docker Compose v2
• a working directory with compose.yaml and a writable .env
• Node.js 24 or later
• the kairos CLI, or willingness to use npx
• Ollama only if the Ollama backend is selected

Remember that a local git clone is optional. The user only needs a controlled copy of compose.yaml plus a writable .env path. Accept either the checked- out repo root compose.yaml or an operator-provided copy in a user-approved working directory.

When something is missing, say exactly what is missing and ask whether to install it now.

Known CLI install command:

npm install -g @debian777/kairos-mcp
kairos --help

Known one-shot npx command:

npx @debian777/kairos-mcp --help

If the user declines a global install, keep using npx @debian777/kairos-mcp <command> for the rest of the session.

Compose v2 verification commands:

docker compose version
docker-compose version

Treat docker compose version as the required success case. If only docker-compose works and docker compose does not, tell the user Compose v2 is missing and ask whether to install or upgrade it.

For Docker, Node.js, or Ollama, use the official vendor installation path for the user's OS after approval if the KAIROS install docs do not already cover the missing prerequisite.

For this skill, assume the KAIROS docs cover CLI and Compose usage, but Docker, Node.js, and Ollama installation steps come from the vendor pages below.

Do not invent your own Docker, Node.js, or Ollama install commands. Use the official OS-specific instructions from the vendor page after the user approves that install path.

If the agent cannot run the official installer directly, stop after opening the vendor page and telling the user what they need to complete manually.

If the OS is not already clear, ask whether the user is on macOS, Windows, or Linux before opening the vendor page.

On Windows, confirm whether the user wants the install to target Windows itself or a WSL environment before you continue.

Success checks after installation:

• Docker: docker --version and docker compose version
• Node.js: node --version
• Ollama: ollama --version or ollama list

Vendor install pages:

• Docker: https://docs.docker.com/get-docker/
• Node.js: https://nodejs.org/en/download
• Ollama: https://ollama.com/download

3. Choose the embedding backend before .env

Explain that the embedding backend determines the .env values and must be chosen before docker compose up.

• OpenAI: managed service; default model text-embedding-3-small; requires a valid OPENAI_API_KEY
• Ollama: local service; recommended model nomic-embed-text; if Ollama or the model is missing, ask whether to install or pull it now
• TEI: existing self-hosted embedding service; ask for TEI_BASE_URL and, if they do not want the repo default Alibaba-NLP/gte-large-en-v1.5, TEI_MODEL

Only use the TEI path when the user already has a reachable TEI service. This skill does not provision TEI for them.

TEI reachability check:

curl -sS "${TEI_BASE_URL}/health"

Treat HTTP 200 from that endpoint as reachable. If the call errors or returns a non-200 response, stop the TEI path and ask the user to fix the TEI service or choose another backend.

Ollama URL rules:

• app in Compose on macOS or Windows, Ollama on host: http://host.docker.internal:11434
• app on the host, Ollama on the same machine: http://127.0.0.1:11434
• app in Compose on Linux, Ollama on host: ask the user for the reachable host IP or published port and do not guess it

Linux Ollama example: if the user says the host-reachable Ollama endpoint is 192.168.1.10:11434, use OPENAI_API_URL=http://192.168.1.10:11434.

Do not continue to .env until the user picks one path.

4. Configure .env

Use the exact environment block from bundled references/docker-compose-simple.md for the chosen backend. If bundled references are unavailable, use the inline fallback snippets below verbatim.

Confirm all of the following before writing:

• target directory
• whether .env already exists
• whether it is safe to create or overwrite the file

Treat overwrite as safe only when either .env does not already exist or the user explicitly approved replacing the existing file after you told them what will change.

Do not merge multiple .env variants in this skill. Create or replace one plain .env file in the chosen working directory.

Treat the directory as writable only when the user approved that location and you can create or replace .env there.

If writing .env would fail with a permission error, stop and ask the user for another directory or for permission changes.

For the simple stack, make sure .env includes QDRANT_API_KEY, AUTH_ENABLED=false, and only the variables required by the chosen embedding backend.

If bundled references are unavailable and you already disclosed the fallback, these are the minimum simple-stack snippets:

OpenAI:

OPENAI_API_KEY=sk-...
QDRANT_API_KEY=change-me
AUTH_ENABLED=false

Ollama with the app in Compose and Ollama on the host:

OPENAI_API_URL=http://host.docker.internal:11434
OPENAI_EMBEDDING_MODEL=nomic-embed-text
OPENAI_API_KEY=ollama
QDRANT_API_KEY=change-me
AUTH_ENABLED=false

If the app runs on the host instead of in Compose, use OPENAI_API_URL=http://127.0.0.1:11434.

TEI:

TEI_BASE_URL=http://your-tei:8080
QDRANT_API_KEY=change-me
AUTH_ENABLED=false

Add TEI_MODEL only when the user wants a model other than the repo default.

TEI with a non-default model:

TEI_BASE_URL=http://your-tei:8080
TEI_MODEL=your-custom-model
QDRANT_API_KEY=change-me
AUTH_ENABLED=false

5. Start and verify

Before starting, confirm that the required ports are free or intentionally remapped.

Default simple-stack ports are:

• app: PORT -> 3000
• Qdrant: 6333 and 6344
• metrics: METRICS_PORT -> 9090

Ask before running:

docker compose -p kairos-mcp up -d

Then verify the server:

curl -sS "http://localhost:${PORT:-3000}/health"

On native Windows outside Git Bash or WSL, substitute the actual port value and use either:

curl.exe http://localhost:3000/health
Invoke-WebRequest http://localhost:3000/health

If PORT=3300 in .env, use http://localhost:3300/health in those Windows commands.

If startup fails, inspect logs:

docker compose -p kairos-mcp logs app-prod

Before using that log command, verify the application service name in the selected compose.yaml. In the repo's default file, the service name is app-prod.

If the service name is different, read the services: keys in compose.yaml and substitute the application service name. Example: if the service key is app, run docker compose -p kairos-mcp logs app.

If you need to check the default ports before starting, inspect listeners on:

lsof -nP -iTCP:3000 -sTCP:LISTEN
lsof -nP -iTCP:6333 -sTCP:LISTEN
lsof -nP -iTCP:6344 -sTCP:LISTEN
lsof -nP -iTCP:9090 -sTCP:LISTEN

If lsof is unavailable, use ss -ltn or ask the user whether those ports are already in use.

On Windows, use:

netstat -ano | findstr :3000
netstat -ano | findstr :6333
netstat -ano | findstr :6344
netstat -ano | findstr :9090

If 3000 or 9090 is busy, ask whether to change PORT or METRICS_PORT in .env before starting. If 6333 or 6344 is busy, ask whether to stop the conflicting service before continuing or explicitly edit compose.yaml.

Concrete conflict examples:

• if 3000 is busy, add PORT=3300 to .env and verify http://localhost:3300/health
• if 9090 is busy, add METRICS_PORT=9191 to .env
• if 6333 or 6344 is busy, either stop the conflicting service or change the host side of the compose.yaml port mappings after explicit approval

If PORT changes, update the MCP URL to the same port. Example: http://localhost:3300/mcp.

If curl fails after up -d, use this order:

1. Run docker compose -p kairos-mcp ps.
2. Inspect the application logs.
3. Re-check .env values for the chosen backend.
4. Confirm the embedding endpoint is reachable.

Failure mapping:

• if docker compose ps shows the app stopped, inspect logs first
• if the app is running but curl still fails, re-check host, port, and /health
• if logs show embedding errors, re-check the selected backend and its URL/key

6. Guide authentication when needed

For the default simple stack with AUTH_ENABLED=false, explain that CLI login is not needed.

If the target server requires auth, guide the user through the CLI flow from bundled references/CLI.md.

Treat auth as required when the user says the server has auth enabled or when kairos token --validate returns 401 or authentication required.

If kairos token --validate fails because the server is unreachable, fix the URL or connectivity first instead of starting the login flow.

If kairos token --validate fails with a server-side 5xx, wait for the server to recover or fix the server health before retrying login.

1. Make sure the CLI points at the correct server URL.

Use either of these patterns:

export KAIROS_API_URL=http://localhost:3000
kairos token --validate

kairos --url http://localhost:3000 search "health"

2. Prefer browser PKCE login when the user can open a browser:

kairos login

3. If the browser must not auto-open, offer:

kairos login --no-browser

4. Validate the stored token when needed:

kairos token --validate

Keep the auth guidance focused on Keycloak-backed CLI usage. Identity-provider administration remains outside this install skill.

7. MCP is optional

Configure mcp.json only if the user's IDE or host explicitly needs MCP over HTTP. The CLI remains primary for verification and operations. Use the example from references/install/README.md#cursor-and-mcp.

Minimum MCP snippet:

{
  "mcpServers": {
    "KAIROS": {
      "type": "streamable-http",
      "url": "http://localhost:3000/mcp"
    }
  }
}

Use this only when the user wants an IDE or another MCP host to connect to the server. Do not create it just because the server exposes /mcp.

If PORT changed from the default, replace 3000 in the snippet with the actual port from .env.

8. Update Docker and npm CLI (maintenance mode)

Run this section when the user asks to refresh Docker and the npm CLI for KAIROS. Keep explicit approvals for each material update action.

1. Confirm update scope: Docker only, npm CLI only, or both.
2. Check current versions first:
   • docker --version
   • docker compose version
   • npm --version
   • kairos --help or npx @debian777/kairos-mcp --help
3. If updates are approved, use official vendor paths for Docker and npm/Node.
4. Refresh KAIROS CLI after toolchain updates:
   • global: npm install -g @debian777/kairos-mcp@latest
   • one-shot: npx @debian777/kairos-mcp@latest --help
5. Validate with the same version and health checks after update.
6. If Docker updates require daemon restart or host reboot, tell the user and stop until they confirm the host is back.

For detailed update commands, cross-platform checks, and fallback handling, use references/docker-npm-cli-updates.md.

9. Install or refresh additional repo skills

Run this section when the user wants more than kairos-install, or asks to sync skills during routine updates.

1. Confirm target skill set:
   • all shipped skills (npx skills add debian777/kairos-mcp)
   • or named skills with repeated --skill
2. List available skills before applying:
   • npx skills add debian777/kairos-mcp --list
3. Install or refresh the selected set:
   • all: npx skills add debian777/kairos-mcp
   • specific: npx skills add debian777/kairos-mcp --skill kairos --skill kairos-bug-report --skill kairos-install
4. If global/agent-scoped install is required, add -y -g and host flags (for example -a cursor or -a claude-code) after approval.
5. Include this skill-sync step when running maintenance updates so Docker/npm and repo skills move together.

Reference modules

Use these reference modules to keep this skill concise:

• references/prerequisites.md
• references/docker-compose-simple.md
• references/docker-compose-full-stack.md
• references/docker-npm-cli-updates.md
• references/examples-and-validation.md

Source: debian777/kairos-mcp — distributed by TomeVault.