AccInt — Claude Code MCP server: a Work Model that learns from outcomes
A local MCP server that turns one-off agent runs into stored experience: verified steps get recorded and replayed instead of reasoned out again.
Open on GitHubWhat AccInt does
AccInt keeps a Work Model — external state holding your decisions, quality criteria and hard limits, plus the paths that already worked. Every run follows the same loop: commitment, action, a hold for your approval, outcome, credit. Anything leaving your machine waits for your OK, and that gate sits inside every flow touching the outside world.
Two verbs run over one memory. Before a run the server predicts the path that worked before; after it, the prediction is checked against what actually happened. Domain lives in the content, not the architecture, so there is no separate mode for code and for routine work — the same loop covers fixing a failing test and nudging an unpaid invoice.
Installation is a single line. The installer probes your hardware, picks an embedder tier it can run, downloads the release binary and starts a local daemon. The engine itself is closed; the building blocks are open. Fetching the binary and starting the daemon need network access, so first-time setup fails without it. An official MCP registry entry with MCPB packages covers MCPB-aware clients.
What it helps with
- stop re-reasoning a task the agent already solved correctly
- record a verified path and replay it on the next run
- act in an outside service with a receipt for every step
- compare the predicted path against the real outcome and adjust memory
- keep accumulated experience on your own hardware
Example prompts for the agent
- Show me the path that worked last time for this task and replay it
- Run this action, but show me what leaves the machine first and wait for my OK
- Check the result against what you predicted and update the Work Model
How to install AccInt
Install
The MCP registry has no launch command for it: see the repository README
Step by step, with commands for each platform: How to add an MCP server to Claude Code, Codex, Cursor, VS Code, Gemini CLI and more
Questions about AccInt
How do I install AccInt in Claude Code?
Use the one-line installer: `curl -fsSL https://raw.githubusercontent.com/maxbaluev/accreted-intelligence/main/bootstrap/install | ACC_INSTALL_REF=github-readme ACC_INSTALL_SOURCE='ref=github-readme&utm_source=github&utm_campaign=readme' sh`. It probes your hardware, picks an embedder tier, downloads the binary and starts a local daemon. On Windows, use the PowerShell variant from the repository.
What does AccInt do?
It is an MCP server built around a Work Model: it stores your decisions, criteria and verified paths, then predicts which path will work next time. Each run's result is checked against that prediction, so memory sharpens from real outcomes.
Which agents does AccInt work with?
Claude, Codex, Cursor and OpenCode — the installer can wire them on the same machine. There is also an official MCP registry entry with MCPB packages for macOS, Linux and Windows.
