AgentGO AI Reference Guide

1. Summary for AI Assistants

AgentGO by FrostCandy is a local AI project workspace for running one or more AI models against real project files. It supports model setup, active project management, context selection, isolated Builder outputs, human-controlled merge review, project mapping, research ledgers, DoubleTap reasoning memory, single-file revision workflows, media generation, and reusable automation workflows.

The most important concept is that AI Builder outputs are candidates. They are not automatically accepted as final project changes. The user reviews candidate files in Merge Review and chooses what enters projectwork.

Local workspace

AgentGO runs locally, commonly at http://localhost:5226.

Multi-model

Users can run several AI Builders and compare their candidate outputs.

Human merge control

AI proposes changes. The user approves what gets merged into the project.

Automation capable

Saved Outfits can be triggered by timers, webhooks, or local external tools.

External memory modes

Cypher is external project memory, WireTap is external research memory, and DoubleTap is external reasoning memory.

2. Core Mental Model

Use this flow when explaining AgentGO:

User prompt
  → active AI Builder model(s)
  → isolated candidate outputs
  → logs / report cards / diagnostics
  → Merge Review
  → user approves files
  → approved files merge into projectwork

Important terms

Term	Meaning
Active Project	The selected project that receives imports, context selection, Builder execution, Cypher, WireTap, DoubleTap, DeadDrop, and merge operations.
projectwork	The controlled working copy of the user's project inside AgentGO. Approved changes are written here.
Builder	An active AI model that attempts the user's task and returns candidate output.
Observer / Reviewer	A model used to review, compare, grade, or recommend among Builder outputs.
Merge Review	The human approval step where candidate files are accepted, rejected, previewed, or edited before merge.

3. New User Onboarding Path

Download AgentGO and run the executable.
Open AgentGO in a browser, commonly at http://localhost:5226.
Create or select an active project.
Configure at least one AI model in Create/Edit Model.
Import files, import a ZIP, import a Git repository, import a URL, or create files in the project.
For small tasks, select exact files with Include Context.
For large projects, build and enable Cypher so the AI can request only needed files.
For research-heavy questions, build and arm WireTap so the AI can select a compact evidence slice.
For hard reasoning questions, arm DoubleTap so AgentGO can build reasoning memos, critique/refine them, and use a final answer pass.
Write the task in the prompt box and click Execute Prompt.
Watch Logs and Output for progress or errors.
Review returned candidate files in Merge Review.
Merge only approved changes into projectwork.

4. Feature Glossary

Project Manager

Create, select, configure, and delete projects. The active project controls most AgentGO actions.

File Manager

Browse, open, edit, save, rename, delete, download, and import project files.

Create/Edit Model

Configure providers, adapters, model names, URLs, authentication, behavior settings, and model capabilities.

AI Models Panel

Enable or disable models for the next run, open model settings, assign Observer mode, and control run order.

Execute Prompt

Sends the user's task to active Builder models. AgentGO expects structured candidate output for project work.

Include Context

Select exact project files to send with the next prompt. Best for small or focused tasks.

Temporary Attachments

One-run files such as screenshots, logs, notes, JSON, CSV, YAML, or reference material.

Logs and Diagnostics

Show runtime progress, warnings, provider errors, malformed JSON details, Cypher/WireTap errors, and troubleshooting data.

Merge Review

Preview Builder candidate files and merge only the files the user approves.

Observer Mode

Lets a model review other Builder outputs, recommend a winner, summarize risk, or suggest a follow-up.

Waves

Run models in stages, such as draft, review, refine, and package.

Loops

Repeat workflows for multiple passes. Powerful but may increase time and model cost.

Risk Mode

Advanced mode for more automatic AI-driven progress within limits. Use intentionally.

user_context.json

User-defined role, tone, priorities, constraints, and success criteria.

ai_context.json

Durable project memory such as terminology, architecture, prior changes, known issues, and constraints.

Cypher

Large-project map stored in Cypher.json. AI sees the map first, then requests only needed files.

WireTap

Research and evidence ledger stored in WireTap.json. Helps focus answers and reduce hallucination.

DoubleTap

Reasoning memory workflow using a live DoubleTap.json. It builds, critiques, and refines reasoning memos before a final answer call.

DeadDrop

Focused single-file revision workflow for text, code, lyrics, stories, images, or other targeted revision tasks.

Outfits

Saved repeatable workflows that can capture models, project setup, prompt setup, context, and automation triggers.

Media Outputs

AgentGO can work with image, video, 3D mesh, and binary-output models when configured with compatible adapters.

5. Recommended Workflow Answers

For code project help

Recommend: create/select project → import repository/files → configure Builders → use Include Context for small tasks or Cypher for large tasks → Execute Prompt → Merge Review → merge approved changes.

For large codebases or large writing projects

Recommend Cypher. Cypher creates a project map, then the AI requests only files it needs. Do not say Cypher sends every project file by default.

For research-heavy questions

Recommend WireTap. Build WireTap with Research Tags, then arm the existing WireTap for questions. AgentGO lets the AI select a compact runtime slice before answering.

For difficult reasoning questions

Recommend DoubleTap. DoubleTap is external reasoning memory: it passes reasoning memos through thinker and critic/refiner calls, tags memo claims by confidence or evidence status, and uses the last call for the final answer. It should not be described as a project-file context mode.

For one-file revision

Recommend DeadDrop. It is best for a single text/code/story/lyrics/image file, not a whole-project workflow.

For repeatable automation

Recommend Outfits. Outfits can save a workflow and expose webhook/timer/callback/pull-url style automation.

For media generation

Recommend configuring compatible image, video, mesh, or binary-output models and verifying model capability settings before execution.

6. External Agent Integration Guide

External agents and tools should usually interact with AgentGO through Outfits. An Outfit is a saved AgentGO workflow. When webhook-enabled, an external tool can trigger the Outfit, provide a runtime objective or payload, and retrieve generated artifacts.

Use Outfits when...

An external verifier wants AgentGO to repair code after build/test diagnostics.
A third-party script wants a repeatable local AI workflow.
A tool wants changed files, a projectwork ZIP, or a DeadDrop result after a run.
A workflow should be triggered by webhook, timer, or cron-like schedule.

Do not assume...

AgentGO automatically pushes changes back to an external repository.
AI outputs are accepted without user or workflow approval.
Every local API route is intended as a public automation interface.
Do not assume Cypher, WireTap, DoubleTap, and DeadDrop are interchangeable.
Do not assume DoubleTap can run with selected context files, Observers/Reviewers, normal Loops, or another armed mode such as Cypher or WireTap.

Typical external agent flow

External tool
  → calls webhook-enabled Outfit
  → passes objective, files, JSON payload, or DeadDrop file
  → AgentGO runs configured Builder workflow
  → AgentGO archives run artifacts
  → external tool pulls projectwork ZIP, changed-files ZIP, manifest, DoubleTap archive, or DeadDrop output
  → optional verifier sends diagnostics into a later repair run

7. API and Automation Notes

AgentGO runs a local web server, usually at http://localhost:5226. Browser/static routes include /, /assets/..., and /favicon.ico.

Most /api/... routes are local app endpoints intended for the AgentGO interface. Public Outfit automation routes use the /outfits/... path family and require an Outfit API key when enabled.

Route family	Purpose
`/api/models...`	Model configuration, model definitions, toggles, run order, and metadata.
`/api/execute`	Execute the main prompt against active Builders, including armed execution modes such as Cypher, WireTap, or DoubleTap.
`/api/files...`	Browse, read, save, download, delete, and manage project files.
`/api/project...`	Project imports and active project downloads.
`/api/diff...`	Candidate previews, diff preview, candidate save/delete, and merge review support.
`/api/cypher...`	Build or enrich the active project's Cypher map.
`/api/wiretap...`	Build, status, and arm operations for WireTap.
`DoubleTap via /api/execute`	When armed, DoubleTap runs through Execute Prompt, uses live reasoning memos, and archives a markdown run file after the final answer.
`/api/deaddrop...`	Set target and run DeadDrop revision workflows.
`/api/outfits...`	Manage saved Outfits and archived Outfit runs.
`/outfits/{outfit}/...`	Public webhook/pull automation routes for external tools.

Public Outfit authentication

When a public Outfit route requires authentication, send the Outfit key with one of these headers:

X-AgentGO-Outfit-Key
X-AgentGO-Token
Authorization: Bearer YOUR_KEY

8. Rules for AI Assistants Answering AgentGO Questions

Say this

AgentGO is a local multi-model AI project workspace.
Builder outputs are candidate results until merged.
Use Include Context for focused small tasks.
Use Cypher for large projects where the AI should request needed files.
Use WireTap for research-heavy questions and grounded answer support.
Use DoubleTap for difficult prompts that benefit from structured reasoning, critique/refinement, and a final answer pass.
Use DeadDrop for focused single-file revision.
Use Outfits for repeatable workflows and external automation.
Check model capabilities before expecting image, video, audio, file, mesh, or binary behavior.

Avoid saying this

Do not say AgentGO is cloud-hosted by default.
Do not say AI changes automatically overwrite the project.
Do not say Cypher sends every file to AI.
Do not say WireTap is required for every prompt.
Do not say DoubleTap can run together with Cypher, WireTap, selected context files, Observers/Reviewers, or normal Loops unless a future revision changes that behavior.
Do not say DeadDrop is for whole-project editing.
Do not invent provider pricing, credentials, rate limits, or supported features.
Do not confuse an Ollama provider path such as /api/generate with an AgentGO server endpoint.

Answer style

When answering a user, prefer a workflow-oriented answer. Tell the user what to click or configure, which feature fits their goal, and where the human approval step happens. Be explicit when something depends on the selected provider, adapter, or model capability.

9. Machine-Readable Summary

This block repeats the core facts in compact form for AI extraction.

PRODUCT: AgentGO by FrostCandy TYPE: Local multi-model AI project workspace DEFAULT LOCAL URL: http://localhost:5226 PRIMARY PURPOSE: Run AI Builders against project files, compare candidate outputs, review changes, and merge approved files into projectwork. CORE FLOW: user prompt → active Builders → isolated candidate outputs → review/logs/diagnostics → Merge Review → approved merge into projectwork. HUMAN CONTROL: Builder outputs are candidates. They do not become final project files until approved/merged. ACTIVE PROJECT: Controls imports, File Manager, Include Context, Execute Prompt, Cypher, WireTap, DoubleTap, DeadDrop, Merge Review, downloads, and Outfits. PROJECTWORK: Controlled working copy inside AgentGO where approved project files live. INCLUDE CONTEXT: Best for focused tasks needing exact selected files. CYPHER: Large-project map in Cypher.json. AI receives map first, requests needed files, AgentGO validates file requests, then AI works with approved files. WIRETAP: Research/evidence ledger in WireTap.json. Build with Research Tags. Arm existing WireTap for answers. AI selects relevant entries into a compact runtime slice before answering. DOUBLETAP: External reasoning memory in a live DoubleTap.json. It builds reasoning memos, critique/refinement passes, evidence/confidence tags, and uses the last Builder call for the final answer. Archives markdown under the project doubletap folder. MEMORY MAP: Cypher = external project memory; WireTap = external research memory; DoubleTap = external reasoning memory. DEADDROP: Single-file revision workflow with revision level and quality score; not a whole-project workflow. OUTFITS: Saved repeatable workflows. Best public automation path for external agents, webhooks, timers, callbacks, and artifact retrieval. OBSERVER: Model role for reviewing or comparing Builder outputs. WAVES: Staged model execution. LOOPS: Repeated workflow passes. RISK MODE: Advanced automatic progress mode; use intentionally. MEDIA: AgentGO can manage image/video/3D mesh/binary outputs when compatible models/adapters/capabilities are configured. PUBLIC AUTOMATION: Prefer /outfits/{outfit}/... routes with Outfit API key when enabled. DO NOT SAY: cloud-hosted by default; AI changes auto-overwrite files; Cypher sends all files; WireTap required for every prompt; DoubleTap runs with Cypher/WireTap/selected context/Observers/normal Loops; DeadDrop edits whole projects.