AccInt

Purpose

AccInt is a local Work Model for enabling AI agents to repeatedly perform development and operational tasks in an accountable, traceable state. It runs beneath the agent tools the user already has (such as Claude Code, Codex, Cursor, OpenCode), turning the agent's behavior into a work model that the user owns: recording commitments, evidence, approval thresholds, outcome credit, and reusable runtimes.

Main Capabilities

• Accountable work state: Every step of execution is written down, forming a record the team can present, while also serving as experience the Work Model can inherit.
• Approval threshold: Any operation that leaves the local machine requires the user's explicit consent, and the user retains control over the real account.
• Outcome credit: Whether a task is effective is judged by real-world results, not merely by the model's self-report.
• Reusable runtime: When the reasoner finds an effective path (browser flow, script, tool integration), it saves it as a runtime; effective flows can be replayed at low cost, while failed flows lose credit and are no longer trusted.
• Local operation: One program, one data file, running on a computer controlled by the user, with no cloud component and no API key required.
• Autonomous evaluation: The built-in evaluation tool maintains recall@5 = 1.000 against exact brute-force retrieval on the live Work Model (based on the material's claim); thresholds, null values, and routing policies are recalibrated based on real results.
• Diagnostic reporting: When a problem is encountered, run acc report to generate desensitized diagnostic information, containing no Work Model content or confidential information, for the user to submit themselves.

How to Use

1. Install AccInt beneath your existing agent tool (supports Claude Code, Codex, Cursor, OpenCode; VS Code, Hermes, and OpenClaw are said to be coming soon).
2. Give the task in ordinary language, and the agent executes it step by step in the real account.
3. Operations that need to leave the local machine will pause and wait for the user's approval.
4. Execution records, approvals, and results are written into the Work Model, and successful paths are saved as reusable runtimes.
5. The user can run acc report to generate desensitized diagnostics and submit them themselves.

The installer detects VRAM, memory, and disk space, chooses the appropriate tier, and prints the reasoning; when hardware is insufficient, it degrades level by level without overstating hardware capabilities. Stronger hardware supports document and screenshot vision features, while ordinary machines run in text mode only.

Applicable Scenarios

• Repeated development or operational tasks, especially scenarios where similar work occurs repeatedly, can reduce repetition costs.
• Teams that need auditing and accountability traceability: multiple operators share one Work Model, with per-person authorization and a unified audit surface.
• Users who are sensitive about data sovereignty and want all work state stored on their own hardware.
• Users already using Claude Code, Codex, Cursor, or OpenCode, to turn agent work into accumulated expertise.

Limitations and Not Yet Provided

• AccInt is at an early stage, and the material explicitly warns that rough edges may be encountered.
• The webpage does not provide specific installation commands, pricing, or billing methods; this information is not yet provided.
• Ideas about a collective Work Model and peer-to-peer cumulative intelligence are future directions; currently there is only one Work Model.
• The specific supported operating systems and minimum hardware requirements are not clearly listed in the material.