Extending the team

Add a team member (no code)

Add an [[agents]] entry to .team.toml:

[[agents]]
name = "Sentinel"
role = "reviewer"
provider = "claude"
model = "claude-sonnet-5-5"
description = "Security reviewer: authentication, secrets, input validation."
owns = ["src/auth/"]

Sentinel now appears in every prompt’s roster, can join discussions (agent-team discuss "..." --members Atlas,Forge,Sentinel), and must sign off any PR that touches src/auth/. A second builder works the same way: agent-team build "..." --builder Anvil, or --parallel to split one plan across every builder. The built-in security, qa and docs roles join build automatically; see configuration.md.

Give a member skills (no code)

skills = ["team/skills/api-style.md", "~/.claude/skills/security-review"] on a member adds those Markdown files (or skill folders with a SKILL.md) to its prompt, on top of its role. Use it for checklists, conventions and domain knowledge one member needs; see configuration.md.

Add a role (no code)

Use any role name and give the member a prompt file:

[[agents]]
name = "Babel"
role = "localisation"
provider = "claude"
model = "claude-sonnet-5-5"
prompt = "team/babel.md"
owns = ["locales/"]

team/babel.md may use {name}, {role}, {project}, {repo}, {roster}, {guarantees}, {owns} and {test_command}.

Add a machine

• Another machine with Ollama: add a provider with kind = "ollama" and its url, for example http://studio.local:11434. Ollama listens only on its own machine by default: on the machine that runs it, set OLLAMA_HOST=0.0.0.0 (on macOS: launchctl setenv OLLAMA_HOST 0.0.0.0, then restart the Ollama app; on Linux: Environment="OLLAMA_HOST=0.0.0.0" in its systemd service). Ollama has no authentication, so only do this on a network you trust.
• A GPU server with vLLM, LM Studio or llama.cpp: kind = "openai" and its url.
• Anything else, or a machine that should keep its own credentials: run agent-team serve there and add kind = "agent-http". See remote-agents.md; agent-team discover finds such nodes on the local network.

Every kind can be the devops member: Claude runs ship on the Anthropic tool runner, and Ollama, OpenAI-compatible servers and remote nodes run it through the same tools with their own tool calling (tool_loop.py). Pick a model that is good at tool use.

Add a provider kind (code)

# src/agent_team/providers/my_provider.py
from .base import Provider, Reply, register, text_messages

@register("my-kind")
class MyProvider(Provider):
    def complete(self, system, messages, *, model, effort, label):
        text = call_my_service(system, text_messages(messages), model)
        return Reply(text=text, content=text)

    def check(self, model):
        return "ok"

Import it in providers/__init__.py, add a test, and kind = "my-kind" becomes valid in .team.toml. Set supports_images = True only if the endpoint accepts Anthropic-style image blocks.

Add a workflow (code)

A workflow is a function run(team: Team, ...) in workflows/. Find members with team.by_role(...) or team.get(name), build them with agent.make(team, cfg, "<prompt>"), exchange files with files.write_files(..., allowed=...), and log with files.Log. Add the command in cli.py and tests with the scripted provider in tests/helpers.py.