Agent runtimes

Claude Code

Claude Code is the primary runtime Agentlas targets. Every draft is shaped to drop into a Claude Code project as-is — no conversion step, no Agentlas-hosted server in between.

What the integration does

Agentlas compiles your one-sentence request into a Claude Code-native project layout — agents, skills, settings, and memory as plain files — instead of a proprietary export format. You review the package on the web, then drop the folder straight into a Claude Code project and run it on the Claude account you already pay for.

Where files go

.claude/agents/<name>.md

The agent file. Frontmatter has name, description, tools.

.claude/skills/<name>/SKILL.md

Each skill is its own folder with a single SKILL.md.

.claude/settings.json

MCP servers and the permissions scaffold.

memory/

Long-running notes the agent can read on each session.

CLAUDE.md

Optional. Top-level project doc. We don't generate it for you yet.

Supported workflows

Single-agent workers — one .claude/agents/<name>.md file for a single recurring job.
Multi-agent teams — a coordinator agent plus specialist agents that hand off work top-down, not peer-to-peer.
MCP-backed agents — auto-detected integrations (Slack, Gmail, Google Calendar, GitHub, Notion, Linear, Supabase) wired into .claude/settings.json.
Imports — paste a public GitHub repo or upload a ZIP of an existing Claude Code project; Agentlas scans it, redacts secrets, and keeps only the manifest.

Example agent team

A request like "triage every new support ticket and draft a calm first reply" compiles to a coordinator (routes each ticket) plus two specialists — a classifier agent and a reply-drafting agent — each shipped as its own .claude/agents/*.md file with a shared memory/ folder for tone and prior-ticket context.

Setup steps

Describe the recurring job in one sentence in Agentlas.
Answer up to 6 clarifying questions (trigger, permissions, model choice, tools).
Review the generated agent or team on the web — what it does, what it can touch, what needs your review.
Download the .zip and unzip it into a Claude Code project, or import it directly if you're already in one.
Open the project in Claude Code. The agents, skills, and settings are already wired — no extra configuration step.

Security model

Every publish runs a 9-category scan (leaked keys, remote-exec payloads, destructive shell, credential exfiltration, broad permissions, and more) before the package is downloadable. Loop guards — no self-invocation, a 3-retry cap on identical calls, a 25-tool-call-per-turn cap, a 5xx circuit breaker — are written into every agent file as instructions Claude Code follows at runtime.

INFO · Full security model

The complete permission, scan, and audit model — including what Agentlas Desktop does differently for local/third-party imports — lives at Docs → Security.

Limitations

Agentlas does not host or proxy Claude API calls. You use your own Claude account and pay Anthropic directly for usage.
There is no live trace viewer yet inside Claude Code sessions — the loop-guard text and the Audit report are today's debugging surface.
CLAUDE.md (the top-level project memory file) is not auto-generated yet; you write or adapt your own.

Common questions

Does this replace Claude Code?

No. Agentlas designs and packages what runs inside Claude Code — the agent files, skills, and settings — it isn't a separate runtime.

Can I edit the generated files by hand?

Yes. Everything is plain markdown and JSON. Nothing is a compiled or obfuscated artifact.

What if I already have a .claude/ project?

Import it as a ZIP or public GitHub repo. Agentlas scans it, redacts secrets, and keeps only the manifest for review — your source isn't retained.

Codex CLI — AGENTS.md is generated alongside .claude/ so Codex-style runtimes read the same worker.
MCP servers — auto-detected Slack, Gmail, Google Calendar, GitHub, Notion, Linear, Supabase, MongoDB integrations.
Agentlas Desktop — the local operations office for running the same package on a schedule without touching a CLI.