From the pack / Product notes
Collar: graph-backed context is not the whole agent
Dotdog prepares queryable project context. Collar is a separate agent harness that consumes graph-backed guidance while coordinating model and tool interactions.
The problem is context assembly
A coding agent needs more than a long system prompt. It needs the right project rules, current artifacts and tool results for the task at hand. Collar organizes behavioral guidance around .dag representations rather than treating every document as undifferentiated prose.
The current project README describes graph-backed system/skill loading, model routing and persistent session state. These are the project's documented capabilities; this article is not an independent reproduction of its model-provider behavior.
How it differs from Dotdog
| Dotdog | Collar |
|---|---|
| Compile/import/query specs and code evidence. | Run the agent conversation and tool workflow. |
| CLI plus local stdio MCP. | Interactive/headless harness with configured model providers. |
| No model account needed for local graph operations. | Provider setup and authorization depend on the chosen backend. |
Use Dotdog with your existing agent host when graph context is all you need. Evaluate Collar when you specifically want its harness behavior. Neither requires turning the other into a hosted platform.
Read the graph, verify the action
A compact representation can reduce an input payload on a particular task, but that does not establish better answers or lower total cost. Model choice, retries, tool calls and output length all affect a run. We are not carrying old compression figures or router timings onto the website without a reproducible, scoped comparison.
Graph text is also data, not blanket execution authority. Follow repository instructions, inspect evidence freshness and keep independent application tests in the loop.
Start with the actual project
Review Collar's installation instructions, supported providers and security guidance before connecting accounts or tools. Keep credentials out of graph exports and published benchmark logs.
Credits
Collar is a specdog project by Justin DiClemente. It is forked from NousResearch/hermes-agent; the upstream work and license attribution remain part of the project.