Introduction
Amodal is a git-repo-centric cloud platform for agents. Your agent definition — connections, skills, knowledge, tools, stores, intents, evals, and optional runtime app code — lives as files in your repository. Amodal reads these files, deploys them to a hosted runtime, and gives your team the operational surface around the agent: sessions, costs, logs, source history, deployments, and evals.
How It Works
Your Repo
├── amodal.json ← agent identity and runtime features
├── amodal/
│ ├── connections/ ← API contracts, docs, and policies
│ ├── skills/ ← Markdown reasoning frameworks
│ ├── knowledge/ ← Domain knowledge documents
│ ├── stores/ ← Typed persistent data schemas
│ ├── tools/ ← Custom tool handlers
│ └── intents/ ← Deterministic typed actions
├── hooks/ ← Guardrail hooks
├── evals/ ← Test cases for agent quality
└── src/ ← Optional custom runtime app
↓ deploy from Amodal
Hosted Runtime
├── Context Compiler ← builds optimized prompts
├── Token Allocator ← manages context window budget
├── Security Layer ← field scrubbing, output guards, action gates
├── Provider Adapter ← model provider selected in Amodal/platform settings
└── Session Manager ← durable scoped sessionsThe Core Loop
Every agent runs the same fundamental cycle:
Explore → query connected systems, load knowledge, gather context
Plan → reason about findings, decide next steps
Execute → call APIs, dispatch sub-agents, present results, learnSimple questions skip planning. Complex questions get the full loop with multi-agent dispatch and skill activation. The runtime matches depth to the question automatically.
Core Concepts
Before embedding an agent in an application, it helps to separate the runtime concepts that often get conflated:
| Concept | Meaning | Example |
|---|---|---|
| Agent | The versioned behavior definition Amodal deploys: knowledge, skills, tools, stores, hooks, connections, evals, and optional app code. | underwriting-agent |
| Session | One conversation or workflow run. This is the chat history, tool-call history, approvals, traces, and token/cost record for a specific interaction. Resume it with session_id. | session_01J... |
| Scope | The application boundary the agent is operating inside. It partitions data, memory, stores, credentials, and operator visibility. | tenant:acme, case:merchant-123, account:42 |
| Scope context | The current facts your app injects for this turn: route, selected record, user role, tenant id, visible page state, or other grounding data. | { "case_id": "merchant-123", "page": "risk-review" } |
| Session type / agent surface | A curated behavior surface that narrows the loaded tools, skills, and instructions for a particular UI or workflow. | request-intake, admin, review |
The most important distinction is:
scope_idtells Amodal where the agent is operating and what data boundary applies.session_idtells Amodal which exact conversation to continue.
You can have multiple sessions in the same scope. For example, a support app might have ten chat sessions inside tenant:acme; all are isolated to Acme's data, but each has its own conversation history. Passing the same scope_id does not automatically mean "resume the last chat." To continue a specific conversation, pass its session_id.
Key Capabilities
| Capability | What It Does |
|---|---|
| Multi-provider | Anthropic Claude, OpenAI, Google Gemini, and other supported providers |
| Git-native config | Everything is a file. Version, diff, review, and deploy your agent like code |
| Connection plugins | Install or define connection packages for the systems your agent should call |
| Security infrastructure | Field scrubbing, output guards, action gates, scope checking, leak detection |
| Evaluation framework | LLM-judged evals, experiments, cost tracking, multi-model comparison |
| Amodal deployments | Commit source changes, deploy them, promote production, and inspect history |
| React SDK | @amodalai/react components and chat widget for embedding |
| Operations UI | Session replay, activity logs, runtime logs, deploy state, and usage analytics |
Next Steps
- Build With A Coding Agent — API-first workflow for Codex, Claude, CI, and advanced builders
- Project Structure — What goes where in your repo
- Build A Support Agent — Create a working tutorial agent with a public demo API
- Source & Deploy In Amodal — Connect source, deploy, promote, and inspect
- Embedding & Multi-tenancy — Put a deployed agent in your product