-
-
Notifications
You must be signed in to change notification settings - Fork 0
Agent Integration
Rune does not compete with coding agents. It empowers them.
Rune is designed to work with any coding agent.
Examples:
- Claude Code
- Cursor
- Codex
- Gemini CLI
- Aider
- Roo Code
- future agents
Rune does not depend on:
- specific APIs
- model providers
- IDEs
The interface is simply:
.rune/
Coding Agent
│
▼
.rune/
│
┌───────────┴───────────┐
│ │
graph.json file summaries
│ │
└───────────┬───────────┘
▼
source code
Source code should be loaded last, not first.
Modern agents often:
- receive a task
- search files
- read large amounts of code
- consume huge numbers of tokens
Humans work differently.
They begin with:
- architecture
- conventions
- dependencies
- mental models
Rune provides these layers.
User asks:
Add Google OAuth.
Without Rune:
User
↓
Agent
↓
Read 100 files
↓
100k tokens
With Rune:
User
↓
Agent
↓
.rune/
↓
graph.json
↓
file summaries
↓
3 files
Less context.
Faster reasoning.
Lower cost.
Example workflow:
Repository
│
├── src/
├── tests/
└── .rune/
Claude Code reads:
spec.md
graph.json
features/
files/
before loading source code.
Benefits:
- reduced token usage
- better architectural understanding
- fewer unnecessary edits
Cursor can leverage:
.rune/spec.md
.rune/conventions.md
.rune/graph.json
to understand:
- architecture
- coding style
- dependencies
before invoking models.
Codex agents can use:
rune context
to identify relevant files.
Example:
rune context "Implement password reset"Output:
{
"related_files": [
"auth.py",
"user.py",
"email.py"
]
}Gemini CLI workflows become:
Task
↓
.rune/
↓
minimal context
↓
source code
instead of:
Task
↓
entire repository
Aider currently relies heavily on selected files.
Rune helps identify which files should be selected.
Example:
rune context "Add subscription support"returns:
billing.py
subscription.py
settings.py
Multi-agent workflows benefit from:
ownership/
sessions/
features/
allowing specialized agents to focus on specific areas.
Rune deliberately avoids:
- SDKs
- RPC protocols
- servers
The contract is:
.rune/
Any agent capable of reading files can consume Rune.
Multiple agents can consume the same context:
Claude Code
Cursor
Codex
Gemini CLI
Aider
All reading:
.rune/
This creates:
- consistent behavior
- shared architecture
- fewer conflicting edits
Possible integrations:
Expose:
rune context
inside the editor.
Automatically prioritize relevant files.
Inject summaries before source code.
Expose:
spec.md
graph.json
features/
through Model Context Protocol.
These integrations belong outside the Rune core.
Initialize:
rune initIndex:
rune indexUpdate after changes:
rune updateRetrieve context:
rune context "Add Google OAuth"Then allow the agent to load source code only when needed.
Long-term architecture:
Human
│
▼
Agent
│
┌─────────┴─────────┐
│ │
MCP Rune
│ │
External Tools .rune/
MCP provides capabilities.
Rune provides understanding.
Agents provide intelligence.
Rune should work with agents that do not exist yet.
Compatibility should come from:
- files
- standards
- simplicity
not APIs.
Agents change.
Models change.
Editors change.
Repository understanding should remain.
Git stores history.
Rune stores understanding.
.git/ explains what changed.
.rune/ explains what the codebase means.
Rune Context — Git for repository understanding.
Git for repository understanding.
- RCP-001 — Repository Format
- RCP-002 — Plugin Protocol
- RCP-003 — Graph Format
- RCP-004 — File Summary Format
- RCP-005 — Feature Map Format
- RCP-006 — Ownership Metadata
- RCP-007 — Session Memory
- RCP-008 — Context Retrieval API
- RCP-009 — Incremental Indexing
- RCP-010 — Multi-Agent Coordination
Git stores history.
Rune stores understanding.