🦌 deer.sh

The AI Elasticsearch Engineer

deer.sh is an AI agent purpose-built for debugging and managing Elasticsearch and the data infrastructure around it. Read-only shell access to your nodes for investigation. Isolated VM sandboxes with replayed data for testing fixes. Ansible playbooks for applying changes to production — reviewed and approved by you.

How It Works

                    Read-Only (direct SSH)          Sandbox (via daemon)
Agent Task  -->  Source Host (investigate)  -->  VM Sandbox (test fixes)  -->  Ansible Playbook  -->  Production
                  - Cluster health               - Full root access            - Human reviews
                  - Index stats                  - Replay Kafka data           - One-click apply
                  - Pipeline configs             - Edit Logstash configs
                  - Shard allocation             - Restart services safely

The agent investigates your cluster through a read-only shell, then spins up an isolated sandbox to test changes against replayed production data. When it finds a fix, it generates an Ansible playbook for you to review before anything touches production.

Features

Feature	Description
Read-Only Investigation	SSH into your nodes with a restricted shell — inspect cluster state, read logs, query services
VM Sandboxes	Full microVM isolation for the agent to test pipeline changes, restart services, experiment freely
Kafka Data Replay	Capture production Kafka topics with PII redaction, replay into Redpanda inside sandboxes
Elasticsearch Sandboxes	Spin up ES inside sandboxes to verify cluster config changes against real data patterns
Ansible Playbook Generation	Auto-generate playbooks from agent work — human reviews before any production apply
Interactive TUI	Natural language interface in your terminal — describe the problem, watch the agent work
MCP Integration	Use deer tools from Claude Code, Cursor, Windsurf
PII Redaction	All tool output scanned for secrets, keys, and connection strings before reaching the agent
Network Isolation	Sandboxes have no route to production — external network access requires explicit approval

The TUI

The TUI is the agent that runs locally on your machine. It has two modes of operation:

Read-Only Mode

The TUI connects directly to your servers over SSH using your existing ~/.ssh/config — no daemon required. A restricted deer-readonly shell ensures the agent can only observe, never modify.

What the agent can do:

Inspect cluster health, index stats, shard allocation (curl localhost:9200/_cluster/health, _cat/indices, _cat/shards)
Read logs, configs, and pipeline definitions (cat, journalctl, tail)
Query services and system state (systemctl status, ps, df, ss)
Run diagnostics (dig, ping, lsblk)

What the agent cannot do:

Write, modify, or delete any files
Start, stop, or restart services
Install packages or execute scripts

Commands are validated twice: client-side against an allowlist in the CLI, and server-side by the restricted shell on the host. You can extend the default allowlist with extra_allowed_commands in your config.

Edit Mode (Sandboxes)

When the agent needs to test a change, it creates an isolated sandbox VM through the daemon. In the sandbox, the agent has full root access — it can modify configs, restart services, install packages, and test fixes against replayed data. Nothing leaves the sandbox without your approval.

Toggle between modes with Shift+Tab.

Preparing a Host

deer source prepare <hostname>

This runs 4 steps on the remote host:

Installs a restricted shell script at /usr/local/bin/deer-readonly-shell
Creates a deer-readonly system user with that shell
Deploys deer's SSH public key to the user's authorized_keys
Restarts sshd

After prepare, the host appears in /hosts as prepared. The CLI generates an ed25519 key pair at ~/.config/deer/keys/ on first run and reuses it for all hosts.

# List prepared hosts
deer source list

The Daemon

The daemon (deer-daemon) runs on the machine that hosts your sandboxes. It manages the full lifecycle of microVM sandboxes — pulling backing images, booting VMs, provisioning services inside them, and tearing them down.

Hypervisor Backends

The daemon auto-detects the best hypervisor backend based on your architecture:

Platform	Backend	Notes
macOS (Apple Silicon)	HVF (Hypervisor.framework)	Native hardware acceleration
Linux (x86_64/ARM64)	KVM	Native hardware acceleration
Any (fallback)	QEMU TCG	Software emulation, slower

Override with the accel config option if auto-detection doesn't match your setup.

How Sandboxes Are Created

Pull backing image — The daemon connects to VM hosts using the deer daemon SSH key and pulls a snapshot of the source VM disk (via libvirt virsh vol-download or Proxmox API)
Create overlay — A qcow2 overlay is created on top of the backing image so the original is never modified
Boot microVM — QEMU boots the VM with a dedicated network interface (TAP device on Linux, socket_vmnet on macOS)
Provision services — Cloud-init installs and configures Redpanda, Elasticsearch, or other services inside the sandbox
Ready — The agent gets full root SSH access to the sandbox via short-lived certificates issued by the daemon's SSH CA

Network Isolation

Sandboxes boot into an isolated network with no route to production by default. If the agent needs to reach an external service (e.g., to verify a fix against a staging API), it requests access and you approve it from the TUI. All network requests outside the sandbox boundary require explicit human approval.

SSH Key Infrastructure

The daemon runs its own SSH CA. It issues short-lived certificates (30-minute TTL) for sandbox access. Certificates embed identity information and enforce security restrictions (no port forwarding, no agent forwarding, no X11 forwarding).

Data Replay

deer.sh can capture data from your production Kafka topics and replay it inside sandbox VMs — giving the agent real data patterns to debug against without exposing production.

How It Works

Configure capture — Tell the daemon which Kafka bootstrap servers and topics to capture from, including auth (SASL/TLS)
Capture & redact — The daemon consumes messages from your Kafka cluster and persists them as JSON segment files. All PII is redacted before storage — the agent never sees raw production data
Provision Redpanda in sandbox — When a sandbox is created with Kafka data sources, the daemon installs and starts a local Redpanda broker inside the VM
Replay — Captured (redacted) records are replayed into the sandbox's Redpanda instance, giving the agent a fully functional data pipeline to test against

This gives the agent a complete feedback loop: it can investigate an issue in read-only mode, spin up a sandbox with replayed data, test a Logstash pipeline fix against real message shapes, verify the output, and generate an Ansible playbook for production.

Demo

Try the hands-on demos:

ES Cluster Red Demo — Boot a 5-node Elasticsearch cluster locally, kill a node to trigger a yellow state, and let the agent diagnose and fix it
Logstash Pipeline Demo — A Kafka → Logstash → Elasticsearch pipeline with a processing bug for the agent to track down

Sensitive Data Redaction

All tool output is scanned for sensitive data before it reaches the AI agent. This prevents accidental exposure of credentials through commands like cat /etc/ssl/private/server.key or kubectl get secret -o yaml.

Type	Examples
PEM private keys	RSA, EC, ED25519, OPENSSH private key blocks
Base64-encoded keys	Output of `cat key.pem \| base64`, Kubernetes secret values
Kubernetes secrets	`tls.key`, `ssh-privatekey`, `private_key`, `secret_key` fields
API keys & tokens	`sk-...`, `key-...`, Bearer tokens, AWS access keys (`AKIA...`)
Connection strings	`postgres://`, `mysql://`, `mongodb://`, `redis://` URIs
IP addresses	IPv4 and IPv6 addresses

Redaction runs at two layers: inline when each tool returns results, and again before the full conversation is sent to the LLM. The agent sees [REDACTED: ...] placeholders instead of actual values.

Skills

The agent ships with built-in skills that give it deep domain knowledge for Elasticsearch and the surrounding data stack. When the agent encounters a problem, it can load a skill to get step-by-step playbooks, common failure modes, and diagnostic commands.

Built-in Skills

Skill	Description
elasticsearch-audit	Enable, configure, and query ES security audit logs
elasticsearch-authn	Authenticate via native, LDAP/AD, SAML, OIDC, JWT, or certificate realms
elasticsearch-authz	Manage RBAC: users, roles, role mappings, document/field-level security
elasticsearch-esql	Query data with ES\|QL, analyze logs, aggregate metrics, build charts
elasticsearch-file-ingest	Ingest CSV/JSON/Parquet files with stream processing and custom transforms
elasticsearch-security-troubleshooting	Diagnose 401/403 failures, TLS problems, expired API keys, role mapping mismatches
kafka	Topic management, consumer group monitoring, cluster health diagnostics
kibana-alerting-rules	Create and manage alerting rules via REST API or Terraform
kibana-audit	Configure Kibana audit logging for saved object access, logins, and space ops
kibana-connectors	Manage connectors for Slack, PagerDuty, Jira, webhooks, and more
kibana-dashboards	Create and manage Kibana Dashboards and Lens visualizations
log-aggregation	ELK Stack deployment, Logstash pipeline building, Filebeat configuration
observability-llm-obs	Monitor LLMs: performance, token/cost, response quality, workflow orchestration
observability-logs-search	Search and filter observability logs using ES\|QL during incidents
observability-service-health	Assess APM service health using SLOs, alerts, throughput, latency, error rate
security-alert-triage	Triage Elastic Security alerts — gather context, classify threats, create cases
security-case-management	Manage SOC cases via the Kibana Cases API
security-detection-rule-management	Create, tune, and manage SIEM and Endpoint detection rules
find-skills	Discover and install new skills from GitHub or local directories

The agent loads skills on-demand via the list_skills and load_skill tools — ask it in natural language and it will pull in the relevant knowledge automatically.

Installing Skills

Install additional skills from GitHub or a local directory:

# From GitHub (owner/repo)
deer skills install elastic/agent-skills//skills/elasticsearch/elasticsearch-security-troubleshooting

# From a local directory
deer skills install ./my-custom-skill

# List installed skills
deer skills list

# Remove a skill
deer skills remove elasticsearch-security-troubleshooting

A skill is just a directory containing a SKILL.md file with YAML frontmatter (name, description, version). Drop it in ~/.config/deer/skills/ or install via the CLI. User-installed skills override built-in skills of the same name.

Quick Start

Install

curl -fsSL https://deer.sh/install.sh | bash

Or with Go:

go install github.com/aspectrr/deer.sh/deer-cli/cmd/deer@latest

Launch the TUI

deer

On first run, onboarding walks you through host setup and LLM API key configuration.

Architecture

                          Direct SSH (read-only)
deer (TUI/MCP)  -------------------------------->  Source Hosts
       |                                              - deer-readonly user
       |                                              - restricted shell
       |                                              - command allowlist
       |
       +--- gRPC :9091 --->  deer-daemon  --->  QEMU microVMs (sandboxes)
                                   |                  - Redpanda (data replay)
                                   |                  - Elasticsearch stubs
                                   +--- control-plane (optional, multi-host)
                                   |
                                   +--- web dashboard

deer-cli: Interactive TUI agent + MCP server. Connects directly to source hosts via SSH for read-only investigation, and to the daemon via gRPC for sandbox operations and data replay.
deer-daemon: Background service managing microVM sandboxes, SSH CA, snapshot pulling, and Kafka data capture/replay.
control-plane (api): Multi-host orchestration, REST API, web dashboard.
web: React dashboard for monitoring and approval.

MCP Integration

Connect Claude Code, Codex, or Cursor to deer via MCP:

{
  "mcpServers": {
    "deer": {
      "command": "deer",
      "args": ["mcp"]
    }
  }
}

17 tools available: create_sandbox, destroy_sandbox, run_command, edit_file, read_file, create_playbook, and more. See the full reference.

TUI Slash Commands

Command	Description
`/vms`	List available VMs
`/sandboxes`	List active sandboxes
`/hosts`	List configured hosts
`/playbooks`	List Ansible playbooks
`/settings`	Open configuration
`/compact`	Compact conversation
`/context`	Show token usage
`/clear`	Clear history
`/help`	Show help

Copy text by dragging and holding Shift.

Development

Prerequisites

mprocs — Multi-process runner for local dev
Go 1.24+
QEMU/KVM — See local setup docs

30-Second Start

git clone https://github.com/aspectrr/deer.sh.git
cd deer.sh
mprocs

Services:

Web UI: http://localhost:5173
API: http://localhost:8080

Project Structure

deer-cli/        # Go - Interactive TUI agent + MCP server
deer-daemon/     # Go - Background microVM sandbox management daemon
api/              # Go - Control plane REST API + gRPC
sdk/              # Python - SDK for the API
web/              # React - Dashboard UI
proto/            # Protobuf definitions

Running Tests

cd deer-cli && make test
cd deer-daemon && make test
cd api && make test
cd web && bun run build

Downloading MicroVM Guest Assets

The live Redpanda guest integration test in deer-daemon/internal/provider/microvm/redpanda_integration_test.go needs three guest assets on the host:

a base Ubuntu cloud image
a matching kernel
a matching initrd

The daemon expects a QCOW2 backing image. Ubuntu publishes the current Noble cloud image as ubuntu-24.04-server-cloudimg-amd64.img, which is suitable for QEMU microVM use even though the filename ends in .img. A simple symlink to a .qcow2 name keeps the local workflow consistent with deer's image store.

Use the helper script from the repository root:

./scripts/download-microvm-assets.sh

That script downloads the image, kernel, initrd, and SHA256SUMS, verifies the checksums, creates the .qcow2 symlink, and prints the DEER_E2E_* environment variables you can use for the live guest test.

To inspect exactly what it would do without downloading anything:

./scripts/download-microvm-assets.sh --dry-run

Running The Live Guest Test With Lima On macOS

On macOS, the most reliable way to run TestProviderIntegration_RedpandaStartsInGuest is inside a Linux Lima VM:

brew install lima
bash ./scripts/run-redpanda-e2e-lima-host.sh --repo-root "$PWD"

Or from deer-daemon/:

cd deer-daemon
make redpanda-e2e-lima

Enterprise

For teams with security and compliance requirements, deer.sh supports:

Encrypted snapshots at rest — Source images encrypted on sandbox hosts with configurable TTL and secure wipe on eviction
Network isolation — Sandboxes boot into isolated networks with no route to production by default, explicit allowlists for service access
RBAC — Control which users and teams can create sandboxes from which source VMs
Audit logging — Full trail of every snapshot pull, sandbox creation, and destruction
Secrets scrubbing — Configurable per source VM: scrub credentials before sandbox creation or keep exact replica for auth debugging
Scoped daemon credentials — Read-only snapshot capability on production hosts, nothing else

If you need these, reach out to Collin to learn more about an enterprise plan.

Contributing

Fork the repository
Create a feature branch
Make changes with tests
Submit a pull request

All contributions must maintain the security model and include appropriate tests.

Reach out on Discord with questions or for access to test VMs.

License

MIT License — see LICENSE for details.

Star History

Made with 🦌 by Collin, Claude & Contributors

Name		Name	Last commit message	Last commit date
Latest commit History 97 Commits
.github		.github
api		api
deer-cli		deer-cli
deer-daemon		deer-daemon
demo		demo
docs		docs
examples		examples
proto		proto
scripts		scripts
sdk		sdk
shared		shared
web		web
.dockerignore		.dockerignore
.gitattributes		.gitattributes
.gitignore		.gitignore
.goreleaser.yaml		.goreleaser.yaml
AGENTS.md		AGENTS.md
CLAUDE.md		CLAUDE.md
LICENSE		LICENSE
README.md		README.md
RELEASING.md		RELEASING.md
SECURITY.md		SECURITY.md
docker-compose.yml		docker-compose.yml
go.work		go.work
go.work.sum		go.work.sum
lefthook.yaml		lefthook.yaml
mprocs.yaml		mprocs.yaml
public-key.asc		public-key.asc

Folders and files

Latest commit

History

Repository files navigation

🦌 deer.sh

The AI Elasticsearch Engineer

How It Works

Features

The TUI

Read-Only Mode

Edit Mode (Sandboxes)

Preparing a Host

The Daemon

Hypervisor Backends

How Sandboxes Are Created

Network Isolation

SSH Key Infrastructure

Data Replay

How It Works

Demo

Sensitive Data Redaction

Skills

Built-in Skills

Installing Skills

Quick Start

Install

Launch the TUI

Architecture

MCP Integration

TUI Slash Commands

Development

Prerequisites

30-Second Start

Project Structure

Running Tests

Downloading MicroVM Guest Assets

Running The Live Guest Test With Lima On macOS

Enterprise

Contributing

License

Star History

About

Topics

Resources

License

Security policy

Uh oh!

Stars

Watchers

Forks

Releases 13

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages