MEGA Code Harness
A coding agent built for agent development. Harness writes agent code, runs eval suites from your repo, and replays any session exactly as recorded, sub-agent and tool calls included.
The problem
Agent code can pass every test and still behave differently on the next run. Behavior has to be measured in execution.
Agents are nondeterministic, so behavior has to be measured by running them and inspecting the trace. With evals and run history outside the coding environment, there's no direct link from a code change to the result it produced.
Teams run their own models, approval rules, and internal policies. A one-size-fits-all agent forces the team to adapt to the tool.
Coding agents need write access to files and a shell, and many send code to external services. For regulated teams, relying on settings and supervision for safety is a liability, and so is code leaving the network.
Our approach
Evaluate and verify your agent where you write its code. Every command runs sandboxed from the first session.
Eval cases live in your repo and run there, including against live agents deployed elsewhere. Evals bypass the cache, so every machine runs them under the same conditions. When a case fails, replay the session exactly as recorded, sub-agent and external tool calls included, and import run history as OpenTelemetry GenAI traces.
Harness adapts to the environment your team already runs, not the other way around. Define modes that set which tools the agent can call and what requires approval, enforce internal policy with hooks that block a call before it executes, protect the files that matter, and connect the model endpoints your organization already uses.
The sandbox is on in every mode. It masks credential directories and secrets such as API keys and tokens, and halts execution if the sandbox itself fails. Protected-file edits always require approval, in any mode, and no project setting can weaken these rules. Everything runs on your machine, and with a local model, your code never leaves it.
Product tour
Harness runs as a single binary in your terminal. Scroll to step through evaluating, replaying, and running your agent safely.
Run the eval suite and see pass/fail per case.
Replay a failed session exactly as recorded.
Switch between manual, plan, and auto modes. The judge flags risky commands, and protected-file edits require approval.
Run in away mode, return to a per-turn list of changed files, and roll back with /undo.
Fits your workflow
Keep your agent framework and observability stack. Run history moves in and out in standard formats.
Usage telemetry (version, platform, install ID, session count) is documented and can be disabled. Prompts, files, and commands are never sent.
Bring your own API keys and switch providers without changing how you work. Harness supports 10+ providers and local models out of the box, and any OpenAI-compatible endpoint with a single config entry.
MEGA Code Harness early access
Request early access to MEGA Code Harness for your agent development team.
