Reference & guides
emisar documentation
Install a runner, author packs, wire in an LLM. The runner and the packs are Apache-2.0 open source, and the full source — control plane included — is on GitHub too.
Runners
Deploy on
The full runner install: every flag, the service it creates, the config file, and the user it runs as.
Run the runner in a container — what it can see, the two shared mechanics, and a sidecar.
One runner per node as a DaemonSet, with the pod spec as your blast-radius dial.
A Nomad system job placing one runner on every client node, with a worked HCL spec.
Enroll ephemeral runners from one reusable key as autoscaling groups — GCP MIG, AWS ASG, Azure VMSS — boot and terminate hosts.
The fleet
What a runner host must reach on the way out — and why nothing has to reach in.
Groups and labels, enrollment keys, pack credentials, updates, and removal.
Go from one runner to a governed fleet, with a phased rollout and a checklist.
Canary a runner release, verify each batch, and keep the last known-good version ready.
Rotate enrollment keys, understand per-runner tokens, and swap the provider credentials packs read.
AI agents
Connect
Wire Claude Code, Cursor, Claude Desktop, and the CLIs in with the emisar-mcp bridge or a raw API key.
Add emisar to Claude.ai as a custom connector — no key to manage.
Add emisar to ChatGPT through Developer mode — no key to manage.
Choose a local sandbox, connect your coding agent to emisar, and understand what remains accessible.
The fleet
One MCP server entry per account, the CLI's stored accounts, and one cloud connector per account.
The agents connected to your account, what each row shows, and how to mint, rotate, and revoke the key behind one.
Re-run the installer per workstation, prove the client relaunched the bridge, and pin a rollback.
Action packs
Choose, inspect, install, configure, and verify a published action pack.
Install the new pack version on a canary, trust the hash, and keep the prior version ready.
Write, validate, install, and trust a pack you maintain.
Run a private registry on GCS, S3, or any static host with packctl.
Action YAML reference: declared args, validation, limits, and redaction.
Operate
Day to day
Select a runner and action, enter its arguments, dispatch it, and read the result.
The run list and filters, every lifecycle status, live output and byte caps, cancellation, and how one dispatch's runs group.
Create, publish, approve, and review a staged procedure built from declared actions.
When it breaks
Start from the symptom: the first check, the page that owns it, and what to send if you need us.
Contain leaked emisar authority, preserve evidence, restore a known-good path, and verify the old authority no longer works.
Compare credential authority, overlap, and revocation consequences before you rotate or contain one.
Govern access
Team & account
Access
Identity concepts
Provider guides
Okta end to end: the OIDC web app, the separate SCIM app, and group-driven roles.
Entra end to end: the app registration with oid, and the provisioning enterprise app.
JumpCloud end to end: one custom application for OIDC sign-in and directory push.
Keycloak sign-in with PKCE, and what its missing outbound SCIM leaves you to do.
Google Workspace sign-in with the locked issuer, and its directory-sync gap.
Reference
Use live MCP tools from the shell, and inspect the server methods, schemas, recovery rules, and errors.
The on-host emisar binary's operator verbs — connect, packs, events, audit, and signing — with their key flags.
Which component owns each decision, what crosses each boundary, and what happens when one drops out.
The v1 compatibility promise, frozen public contracts, and deprecation window.
Trust boundary, searchable audit, hash-chained journal, redaction on egress.
The output, MCP, audit-export, and retention caps emisar enforces — and what happens at each.
The supported product is the hosted emisar control plane today; plans and audit retention are on the pricing page.
No page matches that filter — it reads page titles, descriptions, and groups, not full page text. Try a product word like runner, pack, policy, audit, or SSO.
Guides
Practical guidance for giving AI agents production access without handing them a shell.
July 2026 · 9 min read
How emisar works
How do you give an AI agent access to production without worrying that it might hallucinate its way into dropping a database or taking the site down?
Read the guideJuly 2026 · 7 min read
How to give an AI agent safe access to production
Agents do their best work when nobody is watching over their shoulder. Coding agents get that freedom from a sandbox — production has no sandbox. What teams try instead, where each one cracks, and the division of labor that holds up.
Read the guideJuly 2026 · 8 min read
Prompt injection for ops teams: your logs are prompts now
The moment an AI agent starts reading your production logs, everyone who can write to them can talk to it. How the trick works — and why the defense that holds isn't a smarter prompt but a shorter list of things the agent can do.
Read the guideUse cases
Worked examples of letting an agent operate real infrastructure through declared, gated, audited actions — not a shell.
Incident response: the 33-hour wipe
A CSI driver reformats a live volume; the agent contains it through declared actions and writes the fix to Terraform.
Read it
The fleet-wide 502
Every backend was green, yet one anycast edge threw intermittent 502s; the agent traced it across five layers and stopped the bleed behind approvals.
Read it
Stuck on something?
Open an issue on GitHub or email us. We answer fast.