01 CONTROL
Authority beyond the prompt
The harness governs the tools an agent can use. A model's suggestion does not become permission to execute.
MODEL-AGNOSTIC CODING HARNESS EARLY ALPHA
Run agents
with control.
AI agents can sound convincing even when the work isn't done. Markhor keeps the model's claims separate from the permissions, actions, and evidence behind them.
Built for existing codebases. Designed for local and cloud models.
TASKINSPECT · READ-ONLY
“Explore this repository and explain what it does.”
AGENT RESPONSEModel-generated
Execution recordSeparate from the model's claims
RECORDED05Model requests
03Files inspected
01Invalid request
Tool calls and source observations retained↗
Illustrative display using counts from an actual alpha test. An audit trail is not proof of semantic correctness.
Independent execution control
01 / YOUR MODELS
02 / YOUR REPOSITORIES
03 / YOUR PERMISSIONS
04 / YOUR EVIDENCE
THE PROBLEM
A plausible answer
isn't the same as completed work.
Agents can misunderstand a repository, skip a step, or report success prematurely. Better models help, but the execution layer still needs independent controls and traceable observations.
02 EVIDENCE
What happened, not just what was said
Inspect tool requests, outcomes, source observations, and Run records alongside the agent's response.
03 FLEXIBILITY
Choose the model, keep the controls
Designed around local Ollama inference and configurable provider routes, without binding execution policy to one model.
THE APPROACH
Less guesswork. More visibility.
Markhor is an agent harness, not another language model. It is the layer between what an agent proposes and what the system actually does.
Choose a task
Point Markhor at an existing repository and describe what you want to accomplish.
Set boundaries
Select a model and the level of tool authority appropriate for the task.
Observe execution
Follow model activity, inspected sources, tool outcomes, and recorded events.
Review the record
Evaluate the final response against what the execution evidence actually supports.
BEYOND THE MODEL'S WORDS
An answer is a claim.
A run leaves a record.
Markhor separates model-generated responses from the tool activity behind them. When an agent says it explored a codebase, you can inspect what the harness actually observed.
- Tool activityRequests made, accepted actions, denied operations and outcomes.
- Source observationsWhich files were actually read, not merely listed or mentioned.
- Run outcomeCompletion, interruption and failure distinguished from the model's prose.
MARKHOR / EXECUTION LAYERSREVIEWABLE
01
MODELProposes an actionSuggestions are not permissions
02
MARKHOR HARNESSChecks and executesAuthority, tools and outcomes remain separate
03
EXECUTION RECORDPreserves observationsSource reads, results and terminal state
Conceptual product flow, not a screenshot or a claim that every agent conclusion has been verified.
WHERE WE ARE
Still building.
Not pretending we're finished.
Markhor is an early alpha under active development. The current focus is controlled execution, meaningful evidence, and improving agent reliability on real software projects.
CURRENT STATE EARLY ALPHA
- ✓ Local Ollama model execution
- ✓ Repository inspection and tool records
- ✓ Read-only and governed execution paths
- ✓ Run-level observability and diagnostics
- ↗ Hardening long-running task reliability
Early-alpha functionality is evolving. Markhor does not claim that logging evidence automatically proves an agent's conclusion is correct.
QUICK ANSWERS
Questions worth asking.
Is Markhor another AI model?
No. Markhor is a harness that connects to AI models and governs how agent tasks interact with tools, repositories, and execution records.
Can Markhor run with local models?
Yes. The alpha supports local Ollama inference and has been exercised with small and larger locally hosted models. Capabilities and performance depend on each deployment.
Does an audit record guarantee correctness?
No. It lets you inspect what was requested, executed, and observed. Semantic correctness still requires appropriate verification and judgment.
Can I download Markhor now?
Markhor is currently an early alpha rather than a publicly released product. Public availability and access details will be announced on this site when ready.
THE LONG-TERM DIRECTION
Make agent work
accountable by design.
Different models. Different infrastructure. One independent layer for control and visibility.