Full Specification
The complete AIM protocol specification is maintained in the repository.
Protocol Version
Current version: AIM 1.0
Core Concepts
Progressive Disclosure (4-Tier Model)
AIM uses a tiered loading model to minimize token consumption:
| Tier | Name | Token Cost | When Loaded |
|---|---|---|---|
| 0 | Index | ~5 tokens/capability | Initialization |
| 1 | Schema | ~50-200 tokens | On relevance match |
| 2 | Instruction | Variable | On commitment |
| 3 | Execution | 0 tokens | On dispatch (never in context) |
Transport Agnosticism
AIM is transport-agnostic. Tier 3 dispatch supports:
- MCP: Model Context Protocol servers
- CLI: Command-line tools
- REST: HTTP APIs
- Code: Direct function calls
- Agent: Delegation to other agents
Enforcement Mechanisms
Three enforcement approaches:
- Static: Deterministic tool-based checks (semgrep, tsc, eslint)
- Semantic: LLM-as-judge evaluation with configurable thresholds
- Injected: Rules included in agent context (guidance only)
Composability
Manifests can be composed through:
- Inheritance: Extend base manifests
- Dependencies: Reference other manifests
- Composition strategies:
most_restrictive,last_wins,priority_weighted
Documentation
For the complete specification, see:
- SPECIFICATION.md - Full protocol specification
- JSON Schema - Machine-readable schema