What it does

AccInt, also written acc, is a local tool that sits underneath coding agents such as Claude Code, Codex, OpenCode and Cursor. It keeps a Work Model on your machine: a record of commitments, actions, approvals and outcomes, weighted by whether something outside the model confirmed the result, such as a passing test or an owner's approval. Before a new run it retrieves from that memory and predicts the path most likely to work, and verified steps can be replayed instead of reasoned out again. Anything that sends, publishes, deploys or uses credentials is held until the owner approves, and each step leaves a receipt. The site says the memory is a single SQLite file, with no cloud control plane and no API key needed. It installs with a one-line script for macOS, Linux or Windows, or as an MCPB package from the Official MCP Registry. The engine source is private; the installer, host adapters, docs and plugins are public under Apache-2.0, and the prebuilt binary ships under a draft EULA. The site is candid that parts are young and that the team and collective stages are not started.

Capabilities

Local Work Model stored in a SQLite file on your own machine
Retrieval and action exposed as MCP tools to the agents you already use
Outcome-weighted memory: results confirmed by tests or owner approval count most
Predicts a better path before a run and replays verified steps
Holds external actions such as send, publish and deploy for owner approval, with a receipt for each step
One command wires it into Claude Code, Codex, OpenCode and Cursor

When to use it

Running the same agent-driven job repeatedly and wanting it to improve with each run
Needing an audit trail of what an agent did and who approved it
Keeping agent memory and learning on your own hardware instead of a cloud service