Adapters

Barry runs a session on one of four coding agents, each through its own adapter. The four are Claude Code, Codex, Cursor, and OpenCode. Barry has no default adapter. The identity or the session chooses one.

The capability declarations are in capabilities.ts, and they describe what Barry implements, not everything each vendor’s tool can do. The adapter notes cover the contracts behind them.

Agents

Agent	ID	Session flag	Credential
Claude Code	claude	None. Set it on the identity, or use an endpoint flag	ANTHROPIC_API_KEY from the identity. Without it, Claude Code uses your own login
Codex	codex	--codex	OPENAI_API_KEY from the identity. Without it, the request carries no credential, and Barry’s session API logs a warning
Cursor	cursor	--cursor	Cursor’s own login
OpenCode	opencode	--opencode	OpenCode’s own configuration

Model endpoints

z.ai and Ollama are model endpoints, not agents. They serve models over the Anthropic Messages API, so a session on either one runs the Claude Code agent against a different base URL. Barry refuses an agent flag combined with an endpoint that agent cannot reach.

Endpoint	Key	Flag	Endpoint’s default model	Address and credential
z.ai	zai	--zai	glm-5.2	https://api.z.ai/api/anthropic, with Z_AI_API_KEY
Ollama	ollama	--ollama	qwen3:4b	BARRY_OLLAMA_URL from the session environment, else the port registry. Barry always sends a placeholder key, never your Anthropic credential

A Claude Code session with no endpoint uses Claude Code’s own.

The model catalog is sdk/lib/models.yaml. It is advisory: -m accepts a model ID the catalog does not list, with a warning.

Before an Ollama session starts, Barry checks whether the model makes a real tool call and prints the result. The session starts whatever the result.

Choose an adapter and model

Choose the adapter for every session of an identity, or for one session:

Flag or command	Effect
barry identity agent set <agent>	Set the identity’s default_coding_agent: claude, codex, cursor, or opencode
barry identity agent clear	Remove it
--codex, --cursor, --opencode	Run this session on that agent
--zai, --ollama	Run this session on Claude Code against that endpoint
--model-provider <key>	Choose the endpoint by key, for example ollama
-m <model>, --model <model>	Use this model
barry identity model set <model>	Set the identity’s default_model. --provider <key> also sets default_model_provider, which a model ID the catalog does not list needs
barry identity model clear	Remove it

barry session start and barry session resume both take the agent and endpoint flags. session resume does not take -m.

For example:

Terminal

barry identity agent set cursor
barry --codex
barry --ollama -m qwen3:8b
barry --model-provider ollama
barry identity model set qwen3:8b --provider ollama

Barry resolves each choice in this order:

Choice	Order
Agent	An agent flag, then the identity’s default_coding_agent. An endpoint flag runs the session on Claude Code, and the identity’s agent does not apply
Endpoint	--model-provider, --zai, or --ollama, then the identity’s default_model_provider, then the endpoint the catalog lists for the model. Two different endpoint flags are an error
Model	-m, then the identity’s default_model when the session runs on the identity’s own agent, then the endpoint’s default model

With --prompt, a Codex session runs through Barry’s session API instead of in the terminal.

Capabilities

Capability	Claude Code	Codex	Cursor	OpenCode
Native tool denial	Yes	No. Barry warns, and the tools stay available	Yes, through the session’s .cursor/cli.json. Cursor has no denial for Glob or LS	No. Barry warns, and the tools stay available
Egress sandbox	Yes, on the z.ai endpoint and Claude Code’s own, not Ollama	No	No	No
System prompt	Appended to Claude Code’s own	None. Guidance reaches the agent only as prompt text	Written to AGENTS.md in the session workspace	None. Guidance reaches the agent only as prompt text
Skills	Mounted as a plugin	Not mounted	Mounted as a plugin	Not mounted
Actions as slash commands	Yes	No	Yes	No
Structured output	Native schema. Prompted and validated on Ollama	Native schema	Prompted and validated	Prompted and validated
Resume	Yes	Yes	Yes	No. OpenCode does not report its session ID to Barry
MCP transports	stdio, HTTP, SSE	stdio, HTTP	stdio, HTTP	stdio, HTTP, SSE
Model discovery	No. Ollama: yes, from the daemon	No	Yes	No

A capability that says yes still depends on how the session started. The egress sandbox applies only in sessions that Barry’s session API runs, and native tool denial in a terminal session follows only the bound named with --bound. The bounds reference gives the full enforcement tables.

Codex and OpenCode sessions have a narrower boundary than their bound suggests. Barry’s filtering removes tools it serves, but it cannot remove the agent’s own shell, so bash, file, and network rules do not bind commands the agent runs natively there.

Local models on Ollama can describe a tool call instead of making one. For a model with fewer than 30 billion parameters, or one whose size the daemon does not report, Barry prints a warning before the session starts. A passing preflight check does not guarantee later tool calls.