Emisar
Allows managing Linux hosts through declared action packs, such as checking uptime, with policy enforcement and outbound-only communication.
emisar
Leave the agent working. Keep production authority bounded.
emisar gives MCP-capable agents a catalog of declared infrastructure actions instead of a shell. Policy decides what runs, what waits for a person, and what is denied. A small outbound-only runner checks the action again on the host before it executes anything.
Start with the public pack catalog and let emisar suggest the packs that match a host. Add your own actions without adding another MCP server to every client.
Start with one host
You need an emisar account, a Linux host with
systemd, and sudo. GitHub CLI with gh attestation verify --bundle checks the
release signature. Without GitHub CLI, the installer asks before it continues
with only the checksum. With --yes, it warns and continues. Allow outbound
HTTPS to emisar.dev:443, registry.emisar.dev:443,
tuf-repo-cdn.sigstore.dev:443, and tuf-repo.github.com:443. The last two
serve the public trust roots used to authenticate release checksums. No GitHub
login is required.
GitHub is the optional release fallback. If you want that fallback, also allow
api.github.com:443, github.com:443, and
release-assets.githubusercontent.com:443.
In the console, choose Connect a runner. Copy the generated command; it contains a fresh, single-use enrollment key.
Run it on the host:
curl -fsSL https://emisar.dev/install.sh \ | sudo EMISAR_ENROLLMENT_KEY=emkey-enroll-... bashThe installer authenticates the signed release checksum and verifies the archive against it. It then creates the service, installs host-matched starter packs, and starts the runner.
Confirm the runner is online in the console, then dispatch
linux.uptimewith a reason. You are done when the output appears and the run is present in the audit trail.Open AI agents and connect your client. Remote MCP clients use OAuth; local stdio clients can use the
emisar-mcpbridge and its browser approval flow.
The complete walkthrough, including expected output and troubleshooting, is at
emisar.dev/docs/quickstart. An agent can
perform and certify the setup with the public
install-emisar skill.
Related MCP server: infrastructure-mcp
How an action runs
AI client
| MCP: discover actions, request one with typed arguments
v
emisar control plane
| authenticate, scope, apply policy, wait for approval when required
v
outbound-only runner
| verify pack hash, validate arguments, enforce local limits
v
declared host command
stream redacted output, journal the attempt, update fleet auditThe action pack defines the contract: the executable, argv shape, argument schema, risk, timeout, output limits, redaction, and side-effect description. The model selects from that contract; it does not invent a command line for the runner to execute.
Adding a pack adds capabilities behind the same MCP surface. Operators do not need to deploy another tool server or reconfigure every agent when the catalog changes.
What holds the boundary
The runner opens an outbound TLS WebSocket and exposes no inbound listener. Commands return through that established connection.
Cloud input is limited to declared actions and typed, schema-bounded arguments. The runner rejects unknown actions and arguments.
Packs are content-addressed: a hash of their contents identifies them. The control plane pins the trusted pack hash, and the runner recomputes it from disk before execution. New or changed custom packs wait for trust.
The control plane evaluates runner scope, risk policy, action overrides, standing grants, and conditional approval before dispatch.
On the host, the runner clamps execution options to the pack's limits and runs the declared binary and argv. The runner redacts output before it leaves the host, and Emisar keeps the redacted output in run history.
The control-plane audit includes denied and pending requests. Every runner also writes its execution attempts and local refusals to a hash-chained JSONL journal.
A runner can optionally require bridge-attested dispatch: intent signed by the customer-authorized MCP bridge with an Ed25519 or ECDSA P-256 leaf key. The control plane then cannot invent or widen a permitted call.
Read the exact guarantees, limitations, and threat model in
.agent/kb/specs/security-model.md.
What emisar is not
It is not a sandbox or process isolator. We recommend using one, such as coop.
It is not a generic
execute(command)tool or a replacement for SSH.It does not replace OS least privilege, change management, or configuration management.
It does not make a permitted destructive action harmless. The safety boundary is only as strong as the actions, pack trust, policy, runner configuration, and host permissions in use.
The staging-only shell pack is the explicit break-glass (emergency-only)
exception to the declared-action model. It is critical-risk, default-denied,
never suggested, and should not be installed on production runners.
Find the right surface
Goal | Start here |
Install, upgrade, harden, or diagnose a host | |
Connect Claude, ChatGPT, Cursor, Codex, or another MCP client | |
Inspect or develop the stdio bridge | |
Browse, install, or author action packs | |
Let an agent install emisar, connect a client, or author a pack | |
Review architecture and trust boundaries | |
Review protocol contracts |
|
Contribute to the control plane | |
Review the production GCP infrastructure |
Repository layout
portal/ Elixir/Phoenix control plane, operator console, website, and MCP API
runner/ Go host runner and operator CLI
mcp/ Go stdio-to-HTTP MCP bridge
packs/ Versioned action-pack catalog
skills/ Standalone customer skills for coding agents
infra/ Production Terraform for emisar on Google Cloud
run Root contributor command for development, tests, gates, and operations
dev/ Development Compose topologies, images, configs, and fixtures
tools/ Go implementations behind the contributor command and CI
dist/ Tracked distribution packages plus ignored generated build output
.agent/kb/ Repository architecture, specifications, runbooks, and rulesEach top-level project has its own AGENTS.md with its architecture, security
rules, and verification gate. Run ./run help for the complete list of
contributor commands.
Develop locally
The recommended path needs only Coop and Docker on the host. It installs every tool version the repository pins in the isolated project image:
./run bootstrap # works before Go is installed
coop build # build the pinned project image once
coop run -- ./run setup # sidecars, deps, migrations, browser tooling
coop shell # enter the development boxThen, inside the shell:
./run seed # explicit, idempotent demo data
./run serve # live reload at the URL printed by Coop
# or: ./run serve --iexFor native development, install the exact versions in .tool-versions with
asdf. You also need Git, Coop, Docker, the PostgreSQL client, ShellCheck, jq,
GNU Bash (with read -N), GNU coreutils, Chrome/Chromium, and ImageMagick. On
macOS, Apple's bundled Bash and BSD utilities cannot run all pack regression
checks:
brew install bash coreutils jq
export PATH="$(brew --prefix coreutils)/libexec/gnubin:$(brew --prefix bash)/bin:$PATH"./run setup checks all prerequisites before starting services; ./run doctor reports every detected version and any mismatch to fix. On macOS,
run ./run certs trust once for this workspace after setup.
The fast loop runs Phoenix in the current environment and keeps only PostgreSQL and Keycloak in the workspace-isolated Coop dependency stack.
./run urls prints this workspace's distinct Portal, metrics, Postgres, and
Keycloak URLs. Coop forks inherit the same setup but receive different ports and
volumes. Setup, serve, and reset never apply seeds unless you ask for them.
Use ./run status for a read-only view of the current workspace,
./run logs [db|keycloak] for its exact sidecar logs, and ./run psql for its
development database. Every canonical gate (./run gate <project>) prints its
current phase and elapsed time; a failure names the phase that stopped it.
The root docker-compose.yml remains the slower packaged topology with the
release Portal image, seeded demo data, three runners, MCP, and signing. Start it
with ./run smoke; it serves http://localhost:4010. See
portal/README.md and dev/README.md.
License
This repository is dual-licensed:
runner/,mcp/, andpacks/are open source under the Apache License 2.0. You can inspect, build, package, and operate the on-host components independently.Everything else, including
portal/, is source-available under the Business Source License 1.1. Non-production use is free. Production use is permitted only as needed to operate the Apache-licensed components or the hosted service under the Additional Use Grant. Other production use requires a commercial license. Each version converts to Apache 2.0 on its Change Date.
See contributing, security,
and the CLA. For commercial licensing, contact
licensing@emisar.dev.
This server cannot be deployed
Maintenance
Related MCP Connectors
Security gateway for AI agents: policy, approval, and audited execution, no secrets shared.
Fail-closed policy guardrails for AI agents running kubectl, terraform, helm, and argocd.
Operate Linux, macOS and Windows from your LLM. Every action runs through an auditable allowlist.
Zero-trust gateway for AI agents: score tool calls, verify agent cards, enforce policy, audit.
Related MCP Servers
- AlicenseBqualityAmaintenanceProvides policy-driven, auditable SSH access to server fleets for AI assistants with zero-trust security controls, command whitelisting, and comprehensive audit logging to safely manage infrastructure.1328Apache 2.0
- AlicenseNot gradedqualityAmaintenanceGive Claude Code, Cursor, and other AI agents safe access to your real infrastructure — without giving them raw SSH access.5MIT
- FlicenseAqualityCmaintenanceEnables AI coding agents to run Kubernetes inspection and Terraform plan/apply operations inside ephemeral gVisor-sandboxed jobs with short-lived, narrowly-scoped credentials, while routing destructive changes through a human approval gate.3-
- AlicenseNot gradedqualityAmaintenanceEnables AI tools to securely inspect remote servers by listing directories, reading configuration files, and performing authorized actions through permission controls, human-in-the-loop approvals, and audit logging. It acts as a bastion that never reveals raw SSH credentials or root passwords to the model.1MIT