v0.2.1 · 9 locales · MIT
Coordinate a 9-role software team with AI coding CLIs.
AiCodingAgentTeam is a Go-based orchestrator that dispatches Codex, OpenCode, Claude-Code and DeepSeek-DSH — without holding any API key — to build, review, and ship code under a deterministic quality gate.
9 roles4 protocols0 keys held
aicat \u2014 bash \u00b7 zsh
01/slide \u2014 phase: run
step 0 / 10typing…
Works with your favourite AI coding CLIs
CodexOpenCodeClaude-CodeDeepSeek-DSHgRPCMCPACPA2A
Why teams use it
Core features
Everything you need to orchestrate AI coding without giving up control.
Orchestration vs execution
The Coordinator orchestrates and never writes code. Host CLIs execute and never decide. No coupling, no blurred responsibility.
Container-as-role
9 team roles — PM, Architect, Frontend, Backend, QA, Security, DevOps — each in its own container, independently scalable and replaceable.
A2A protocol collaboration
Structured inter-agent messaging. InProc mode for development, Redis Pub/Sub for containerized deployment.
Deterministic quality gate
golangci-lint, go vet and go test run as a hard machine check. No relying on the model to grade itself.
RAG + project memory
BM25 retrieval injects context. Memory accumulates failure lessons and historical solutions across runs.
Four external protocols
gRPC for the TUI, MCP for external tools, ACP for standard Agent clients, A2A for cross-instance mesh.
Local-first, keyless
Code stays inside the container. No API keys held. All authentication is delegated to the underlying host CLI.
Scale-adaptive flow
A quick edit skips the DAG. A large project gets the full team. The platform adapts to the work, not the other way around.
// metrics
By the numbers
Designed to be small at the core and big where it matters.
9team roles0x0
4external protocols0x1
9UI languages0x2
0API keys held0x3
$ aicat run
One command, full pipeline
From a natural-language requirement to a verified proof-pack in a single CLI invocation.
aicat run
bash$ ./bin/aicat run "Build a REST API" --backend codex[router] intent=build_api workflow=full_team backend=codex[planner] nodes=9 writers=3 reviewers=6 dag=built (12ms)[sched] dispatched reviewer:qa + reviewer:security + reviewer:arch (parallel)[sched] dispatched writer:backend → writer:frontend → writer:devops (serial)→ 3 writers · 6 reviewers · 1 quality gate · 1 proof-pack✓ proof-pack.zip plan.json · verify.jsonl · scorecard.md · delivery-summary.mdgate: 0 lint · 0 vet · 0 test · score=98 / 100
// flow
Five layers, one flow
Each layer is a small, replaceable component. The interface between them is the only contract.
1
RouterIntent → workflow
Intent router classifies the user's request and chooses the workflow.
layer.01
2
PlannerDAG of writers + reviewers
Planner builds a DAG of writer and reviewer nodes.
layer.02
3
SchedulerParallel reviewers · serial writers
Scheduler dispatches reviewer nodes in parallel; writer nodes serially under a single-writer lock.
layer.03
4
KnowledgeBM25 RAG + project memory
Knowledge (BM25 RAG) and Memory inject relevant context into each node.
layer.04
5
Quality Gategolangci-lint · vet · test (machine)
Quality gate runs machine-checkable verifications; failures surface with structured detail.
layer.05
// 4 steps
From zero to first pipeline
Single-binary mode is the fastest path; TUI and containers are optional layers.
1
1. Build the binary
step.01Compile from source — pure Go, no runtime dependencies.
$ go build -o bin/aicat ./cmd/aicat2
2. Initialize a project
step.02Scaffold the workspace configuration in the current directory.
$ ./bin/aicat init3
3. Run a full pipeline
step.03Pass a natural-language requirement and pick a backend (Codex is the real driver).
$ ./bin/aicat run "Build a REST API" --backend codex4
4. Verify quality
step.04Run the deterministic quality gate on the artifacts produced.
$ ./bin/aicat verify./aicat --ready
Ship a 9-role team in 60 seconds
Pull the binary, write one command, get a deterministic verification artifact. That's the whole pitch.