Design the workflow.
Scale the unicorn.
Colony gives you a visual builder to design exactly how your AI agents work together. Drag nodes, pick models, set rules. Then watch your swarm execute—live, in parallel, across your entire codebase.
STEP 01 — DESIGN
Build your workflow like you'd build a circuit.
Open the Agent Builder. Drag nodes onto the canvas: plan, code, test, review, deploy—whatever your workflow needs. Connect them. Each node is a step your AI swarm executes. You decide the order, the branching, the gates. No fixed pipeline. No assumptions about how you work.
9 Node Types
Start, Plan, Subtasks, Code, Test, LLM Review, Human Review, Deploy, End. Combine them however you want.
Per-Node AI Selection
Claude Opus for planning. Codex for code. Gemini for review. Pick the right model for each step.
Custom System Prompts
Override the default instructions on any node. Your workflow, your rules.
STEP 02 — ASSIGN
Different models. Different jobs. One workflow.
Each node in your workflow can run on a different AI provider. Claude writes the plan. Codex writes the code. Gemini reviews it. Bring your own API keys (BYOK) for total cost control and no markup on token usage.
Claude
Opus 4.6, Sonnet 4.6, Haiku 4.5. Extended thinking for architecture decisions.
Codex
GPT-5.3, 5.2, 5.1. Max and Mini variants for different cost/speed tradeoffs.
Gemini
3 Pro, 3 Flash, 2.5 Pro, 2.5 Flash. Lightweight nodes where speed beats depth.
STEP 03 — VERIFY
One AI writes. Another AI checks.
Cross-AI review with iterative refinement.
Iterative Refinement
Reviewer finds a problem, sends feedback, code node regenerates. Up to 4 cycles.
Auto-Fix Mode
Reviewer can fix issues directly without bouncing back. Faster for small corrections.
Block on Fail
If review fails after all iterations, workflow halts. No bad code slips through.
One model's blind spot is another model's strength.
Evidence Panel
Full diffs, streaming logs, artifact trail. Review with data, not faith.
One-Click Intervention
Approve, reject, or modify at any gate. No terminal commands.
Flexible Placement
Gate after planning, after coding, before deploy. Your call.
STEP 04 — CONTROL
Pause anywhere. Approve anything.
Drop a Human Review node wherever you need a checkpoint. Workflow execution pauses. You see the evidence: diffs, test results, review verdicts, risk assessments. Approve to continue. Reject to stop. The merge decision always lands on a human when you want it to.
STEP 05 — RUN
Hit run. Watch the swarm work.
Your workflow compiles into a directed graph. Colony validates the structure, checks for cycles, ensures single entry and exit. Then it executes—node by node, with real-time streaming logs. Each node's output feeds into the next. Progress updates push to your dashboard live.
Live Streaming
Watch each agent's output as it generates. Not after. During.
State Persistence
Pause and resume from any node. Pick up where you left off.
Error Classification
Timeout, rate limit, permission denied, merge conflict. Every failure has a type.
Start from a template. Customize from there.
Pre-wired workflows for the patterns you run every day.
Feature
Bugfix
Refactor
Or build your own.
Every template is a starting point. Fork it, rewire it, add nodes, remove gates. The Agent Builder is yours.
INTEGRATION
From issue to PR. Bi-directional.
Create a GitHub issue. Colony picks up the workflow. Agents execute.
Labels update in real-time: planning, in-progress, review, done.
When the workflow finishes, a PR is ready.
Build your first workflow in 5 minutes.
Free to start. Your machine. Your keys.
No credit card. No signup. Just download.