Agents

Runtimes

How LoCrew drives Claude Code, OpenAI Codex, models from AI providers and external Agent2Agent agents, and how they differ in sessions, tools, file access, cost reporting and verification.

A runtime is what does an agent's thinking. LoCrew supports four. They share one conversation model, but they differ in what they can do and what they report back.

#Comparison

	Claude Code	OpenAI Codex	AI model	External agent (A2A)
Runs as	Your `claude` CLI via the Claude Agent SDK	Bundled Codex CLI via the Codex SDK	LoCrew's tool loop over a provider's API	A remote agent over A2A
Credentials	Your Claude Code login	Your Codex login	The provider's API key, or none for local servers	None, bearer token or custom header
Working directory	Required	Required	None	None
File access levels	Read only / Ask first / Full access	Mapped onto Codex's sandbox	Always read only	None
LoCrew's diff approval dialog	Yes, tested against the real CLI	Not shown by LoCrew (see below)	Not applicable	Not applicable
LoCrew tools (hand-offs)	Yes	Yes	Yes, if the model accepts tools	No
MCP tools you grant	Yes	Yes	Yes, if the model accepts tools	No
Skills	Yes	Yes	No	No
Memory between messages	Native session per agent and conversation	Thread per agent and conversation	None: rebuilt from the transcript every run	Remote context id per agent and conversation
Shows reasoning	Yes	Yes	If the provider sends it	No
Reports tool calls	Yes	Yes (shell, file changes, MCP)	Yes	No
Cost	Real dollars from the CLI	Estimated from a price table	Not priced (tokens only)	Not reported
Spend ceiling enforced	Yes	Yes (estimate)	No	No
Turn limit enforced	Yes	No evidence in the code	Yes	Not applicable
Verification	Real CLI in an opt-in live suite	Install detection in the default suite; opt-in live tests exist	Mock provider servers	Official A2A SDK server

#Claude Code

LoCrew drives your own claude binary through the official Claude Agent SDK. Your existing login pays for the run; LoCrew never asks for or handles an Anthropic API key.

Detection: claude --version on your PATH. If it is missing: Claude Code was not found on PATH. Install it with "npm i -g @anthropic-ai/claude-code", then reopen this window.
Session: each agent keeps a native Claude Code session per conversation and resumes it on the next message.
Isolation from your personal setup: agents run without your ~/.claude settings, hooks, MCP servers or CLAUDE.md. Tools and skills come only from what you grant in LoCrew.
File access: Read only denies Write, Edit, MultiEdit, NotebookEdit and Bash. Ask first lets reads through and stops every write or shell command for your approval. Full access allows everything.
Limits: the agent's turn limit and spend ceiling are passed to the SDK. A run that hits them ends with The agent reached its turn limit for this execution. or The agent reached its spend limit for this execution.
Auto-compact: the editor's Claude Code section has Let it compact its own history. When the context window fills, the runtime summarises older turns. LoCrew's transcript is never touched; you see a line such as Agent compacted its own history: … to … tokens.
Plugins: if a supported plugin is installed (currently the ponytail plugin), it appears in the same section as a toggle.

#OpenAI Codex

LoCrew ships the Codex CLI through its @openai/codex-sdk dependency and falls back to a codex on your PATH. Sign in with npx codex from the LoCrew folder. The README marks Codex agents as beta.

Session: a Codex thread per agent and conversation, resumed on the next message.
Instructions: Codex has no system-prompt channel, so the agent's persona and instructions are sent at the start of a new thread only.
File access maps onto Codex's own sandbox: Read only → read-only; Full access → workspace-write with no approvals; Ask first → workspace-write with Codex's on-request approval policy.
Cost: estimated from token counts and a built-in price table, and checked against the spend ceiling: The agent reached its spend limit for this execution (estimated $…).

What is not verified for Codex

The default automated suite only detects the Codex install. The project's opt-in live suite includes Codex turns, but whether it has been run is not recorded. The Codex adapter does not raise LoCrew's own approval dialog, so do not rely on Ask first to show you a diff for Codex writes; use Read only or a separate copy of your project. The code also shows no enforcement of Max turns per message for Codex.

#AI model agents

A model agent runs on any model from a provider you connected in Settings → AI Providers. LoCrew runs the tool loop itself.

No files: model agents have no working directory and are always read only. They answer, hand off work and use the MCP tools you grant.
No session: every run rebuilds the context from the conversation's recent messages (up to 60, trimmed to half the model's context window, oldest first). Other speakers appear as [name → you].
Learning capabilities: if a model rejects tools, LoCrew retries without them and remembers: … does not accept tools, so … will answer without them. The same happens for images.
Turn limit: reaching it adds (… reached its limit of N model turns for this message.)
Cost: tokens are recorded, but no price. Spend ceilings do not constrain model agents, so a paid API is not protected by them. The hop and turn limits still apply.

#External agents (A2A)

An agent that runs elsewhere and speaks the Agent2Agent protocol (v1.0 or v0.3, over JSON-RPC or HTTP+JSON). The README marks it beta.

No tools: it cannot call LoCrew's tools, hand work to other agents, or use MCP servers.
Memory: LoCrew stores the remote context id per conversation. Instructions go with the first message of each conversation.
Endpoint pinning: the endpoint you approved is stored. If the card later stops listing it, runs fail with Its Agent Card no longer lists the approved endpoint … until you test the connection again in the agent's settings.
Input or authorisation needed: shown as Waiting for you with ❓. Your next reply continues the same remote task.
Images are sent only if the card accepts them.
Usage: not reported; the usage chip reads Usage not reported.

#Signing in again

When a Claude Code or Codex run fails because the login expired or was revoked, every agent on that runtime is marked:

Signed out of Runtime: the login expired or was revoked. Sign in again to continue.

A banner appears at the top of the window with **Sign in to Runtime and I’ve signed in. Sign in opens a terminal running claude auth login, or codex logout and then codex login. Finish in the browser it opens, then press I’ve signed in. If the check passes, the agents return to Ready**; the next run confirms it.

Note

Codex can report "Logged in" with a revoked token, so only a real run proves the login works.