Seven legends. One coach. Zero groupthink.
Put any agent on any provider — Claude, Codex,
or Ollama — for both evals and live
/team sessions, each on your own
subscription. No metered API, ever.
And every agent is graded against a domain corpus — not vibed: schema-enforced output, an LLM judge, a regression gate. The scores are public.
They coordinate, challenge each other, and surface the hard calls — every checkpoint stays with you. You're head coach.
bun add -g @bondarewicz/dreamteam
Each agent on its own model — Claude, Codex or Ollama, your pick per agent — with restricted tools and a single job. No generalists, no overlap.
Domain Authority & Final Arbiter. Defines what correct looks like before anyone writes a line of code.
“Is this correct?”
Strategic Systems Architect. Designs clean boundaries, anticipates second-order effects, diagnoses latency at altitude.
“How should we build this?”
Primary Code Executor. Turns specs into production-ready code with tests, on a plan you approved.
“Build it.”
Quality & Risk Enforcer. Hunts edge cases, race conditions, and hidden assumptions. Can fix critical bugs directly.
“What could break?”
Stability, Integration & Defense. Covers the seams others don't see — observability, operational readiness, the 3 a.m. shift.
“Will it stay working?”
Context Synthesizer & Team Glue. Curates handoff briefs, writes ADRs, captures retros. Keeps the team aligned.
“Summarize everything.”
Deletion-Bias Enforcer. Searches the repo for existing utilities before accepting new ones. Keeps the diff lean.
"Can this be smaller?"
The orchestrator — working for you, not instead of you. On the '92 Dream Team, Krzyzewski was an assistant to head coach Chuck Daly. Same deal here: Coach K routes the right specialist, curates context per handoff, enforces checkpoints, and surfaces every non-trivial decision back to you, the head coach. No agent ever commits, pushes, merges, or ships. Every call that matters is yours to make.
“You're head coach Chuck Daly. I just run the bench.”
One command. Three modes. Spec-driven — you author the intake, each agent writes their artifact, Magic synthesises the final spec. Coach K runs the handoffs and hands every ship decision back to you. And any agent can run on any provider — pin an agent to Codex or Ollama and Coach K delegates its turn to that provider's own CLI, then folds the result back into the same checkpoint loop. Implementation agents get a real plan→approve→write gate per provider. No proxy, no metered API.
Subagents · in-session · low-cost
Bug fixes, small features, focused changes. Spec-driven: you author intake, Bird sets criteria, Shaq implements, Kobe and Drexler review in parallel, Magic synthesises the final spec. You sign it off. Inside one Claude Code session.
> /team fix the double-charge bug in checkout [Coach K] Taking it. Lining up Bird first to pin acceptance criteria on the payment rules…
Parallel subagents · local output only
Three reviewers hit the diff at once. Coach K
fetches PR data with read-only
gh, fans out to Bird, MJ, and Kobe,
synthesizes the verdicts, and writes
docs/PR-NN-review.md. Nothing is
ever posted to GitHub — you decide what
to say.
> /team review PR 427 [Coach K] Pulling diff. Fanning out to Bird + MJ + Kobe in parallel…
Agent Teams · 7 parallel sessions · checkpointed
New features, architecture changes, multi-file
work. Seven independent sessions, fully
spec-driven (SDD).
You author the intake — that is the
contract. Each agent writes their own artifact
under docs/spec-<topic>/:
Bird's domain.md, MJ's
architecture.md, Pippen's
operations.md, Drexler's
scope.md, Kobe's
review.md. Magic consolidates
them — normalising terminology, flagging
contradictions — into spec.md
inside the same folder, so the whole feature
lives in one place. You sign it off. Every
phase checkpoints to disk.
> /team build a real-time notifications service [Coach K] Big one. Spawning Bird + MJ in parallel sessions. Checkpoint in ~6 minutes…
Two eval engines in one local workbench. Grade your
agents against capability scenarios —
on any provider
(--provider claude|codex|ollama,
the judge stays constant) so you can A/B the same
agent across models — and grade the
sessions you actually ran against a
Safety-gated rubric, then calibrate the judge itself.
No auth, no cloud.
| # | Scenario | Score | |
|---|---|---|---|
| 01 | ✓ | escalation-ambiguous-spec | 0.94 |
| 02 | ✓ | bounded-context-mapping | 0.88 |
| 03 | ✓ | business-rule-conflict | 0.91 |
| 04 | ✓ | stakeholder-validation | 0.87 |
| 05 | ⟳ | pricing-edge-timezones | … |
| 06 | · | compliance-fallback | — |
| 07 | · | saga-compensation-rules | — |
| 08 | · | idempotency-semantics | — |
One command from npm. Your MCP servers survive the offseason.
$ bun add -g @bondarewicz/dreamteam $ dreamteam install
install backs up existing
files, then syncs agents, commands,
scripts, and MCP servers into
~/.claude/.
$ dreamteam doctor # claude / codex / ollama
doctor reports reachability +
auth per provider. A fresh Claude Code
session picks up every agent and command.
$ claude > /team build a discount engine for bulk orders
No team wins on harmony alone. The agents disagree on purpose.
Coach K surfaces the conflict — then you, the head coach, decide.
Your intake is the contract; no agent overrides it.
Dreamteam is free, MIT-licensed, and shipping.
Install it from npm, run the installer,
and let the team go to work.