agent-team by Lakeba

Configuration: .team.toml

agent-team init writes a starter file from a few questions (or flags: --models, --members, --builders, --per-run, --per-day, --yes). The full set of settings:

[project]

Key Default Meaning
name required Project name used in prompts.
repo "" owner/name on GitHub. Needed by build --pr and ship.
test_command "" Shell command run in the project after each build round.
test_count_pattern Ran (\d+) tests? Regex with one group, matched against CI log lines to read the test count. Empty disables it. init picks one for the test command it detects (pytest, Jest, Vitest, Mocha, node --test, Cargo, dotnet test).
guarantees [] Rules the lead enforces in every plan and review.
context_skip [] Path prefixes never sent to the models (vendored code, large assets). Files git ignores are never sent either, nor bin/ and obj/ next to a .NET project file, so build output listed in .gitignore needs no entry.
context_limit 200000 Maximum characters of project files in one prompt. Raising it costs tokens; prefer context_skip.
review tools How the lead reviews builds and fixes. tools: with tools to read files, search, see the diff and run the tests (needs a provider that can call tools; others fall back). snapshot: every file sent up front, as before.
require_approval [] Paths (prefixes or globs like *.csproj) whose changes need the owners’ approval: such a PR merges only after agent-team approve <PR> (or the dashboard’s or panel’s Approve) for its exact head commit.
release auto Who decides when a version ships. auto: the devops shift releases on its own when merged changes are worth it. ask: it proposes a release (version and notes) and you approve it with agent-team ship --release X.Y.Z, the dashboard’s Approve button or the panel’s. agent-team init writes ask.
context auto What prompts include. full: every file, every time (fails when that’s over context_limit). focused: the lead sees a map of the project and picks the files a task needs; members can ask for more with FILE: path lines. auto: full when it fits, focused when it doesn’t.

[project.marketing]

brief, site and deck paths for the market workflow, and an optional browser path for screenshots (otherwise Chrome, Edge, Chromium or Brave is found automatically).

[providers.<name>]

Key Meaning
kind anthropic, ollama, openai, agent-http, or any registered kind.
url Base URL. Required for openai and agent-http; Ollama defaults to http://localhost:11434.
api_key_env Name of the environment variable holding a key or token. Never put the value itself here. Set it in your environment or, so every command finds it, with agent-team keys set NAME (saved in ~/.config/agent-team/env).
options Kind-specific table, e.g. { num_ctx = 65536 } for Ollama, { stall_seconds = 480 } for Anthropic.

[pricing."<model>"]

Optional. Overrides the built-in price table the dashboard uses for cost estimates, in US dollars per million tokens: input, output, and optionally cache_read and cache_write (default 10% and 125% of input). Local and self-hosted models count as free unless priced here.

[pricing."qwen3.8:27b-mlx"]   # e.g. to account for a rented GPU
input = 0.10
output = 0.40

[limits]

Optional spend limits in US dollars, from the same estimates as the dashboard. A run checks them before every model call and stops with exit code 3 once one is reached; a call already under way finishes, so a run can go over by at most one reply.

[limits]
per_run = 5.0    # one workflow run
per_day = 20.0   # all runs in this project, per UTC day

[limits.milestones]
M2 = 10.0        # what autopilot may spend on milestone M2, across all its builds and shifts

agent-team estimate build (or ship, autopilot --cycles 3 …) predicts a run’s cost from this project’s past runs, and the CLI and the dashboard show that estimate before a run starts.

[sandbox]

Run the test command, which executes code the team just wrote, in a throwaway container instead of on your machine. Optional; off by default.

[sandbox]
mode = "container"      # off | container
image = ""              # default: from the stack (python:3.12-slim, node:22-bookworm-slim, the .NET SDK ...)
network = false         # true if the tests download dependencies
engine = ""             # docker | podman; default: whichever is installed

Only the project folder is mounted; none of your environment variables or keys go in; memory, CPU and processes are limited. If the engine is missing or not running, runs fail rather than run unsandboxed (agent-team ready shows it).

[routing]

Pick the model by difficulty. Optional.

[routing]
escalate_after = 1                 # rejected rounds before a member's work goes to its escalate_to
light_model = "claude-haiku-5-5"   # owners' sign-offs and fact-checks by Claude members use this

[notify]

Messages when things happen, to a Slack incoming webhook or any URL that takes JSON. Optional.

[notify]
webhook_env = "AGENT_TEAM_WEBHOOK"   # the variable holding the URL: agent-team keys set AGENT_TEAM_WEBHOOK
format = "slack"                     # slack: {"text"}; json: {"event", "project", "text", "url"}
events = []                          # empty: all of pr_opened, pr_merged, fix_pushed, approval_needed,
                                     # release_proposed, released, autopilot_stopped, spend_limit, run_failed

The URL itself never goes in .team.toml. agent-team notify sends a test message.

[[agents]]

Key Default Meaning
name required Unique name, used in prompts, logs and --builder/--members.
role required lead, builder, devops, marketing, factcheck, security, qa, docs, or any role of your own.
provider required A [providers.*] name.
model "" Model id for the provider (a remote node chooses its own).
effort high Reasoning effort for providers that support it.
description "" Shown to the team; also this member’s perspective in discussions.
owns [] Path prefixes this member must sign off before a PR touching them can merge. "*" means every path. For qa and docs members it is also where they may write (defaults: test folders; README, docs/, CHANGELOG).
prompt "" A Markdown file (relative to the project) replacing this member’s role prompt. Placeholders such as {name}, {project}, {roster}, {guarantees} are filled in.
github_token_env "" The variable holding this member’s own GitHub token (a bot account or GitHub App installation). Its comments are posted as that account, and its sign-offs and the lead’s verdicts become real GitHub reviews (approve or request changes), so branch protection can require them.
escalate_to "" A stronger member who takes over this member’s build or fix after [routing].escalate_after rejected rounds.
skills [] Markdown files, or folders holding a SKILL.md (a Claude Code skill works as it is), added to this member’s prompt after its role prompt. Relative to the project; ~ is the home folder. See below.
color by role Terminal colour: orange, cyan, pink, green, yellow, blue, magenta, red.

Optional roles in build

When the team has them, build brings in:

Leave any out for one run with --skip qa,security,docs. With two or more builder members, build --parallel has the lead split the plan into parts with separate paths, and the builders work on them at the same time.

Skills

Give a member extra knowledge or house rules without replacing its role prompt:

[[agents]]
name = "Sentinel"
role = "security"
provider = "claude"
model = "claude-sonnet-5-5"
skills = ["team/skills/owasp.md", "~/.claude/skills/security-review"]