Why it matters
This tool addresses a critical need for reliability in AI agent workflows by providing deterministic enforcement of action integrity. Builders can use Makoto to prevent agents from acting on assumptions or stale data, leading to more trustworthy and predictable agent behavior.

What changed

Makoto is a recently released Claude Code hook designed to enhance the reliability of AI agent actions. It operates by intercepting agent steps, such as writing, editing, or committing code, before they are executed. Makoto then scrutinizes the session's logged record to ensure that any information required for the step is present and accurate according to its defined principles of "Immediacy." These principles are:

  • Original: Prioritize reading original data over derivations.
  • Unchanged: Do not assume that data which can change has not.

If a step requires information that is missing or violates these immediacy rules, Makoto will block the action. For instance, if an agent claims to have tested code but the session record doesn't show a corresponding test execution, Makoto will prevent the step. The checker itself is deterministic, using no external models or network calls, and aims for zero false positives in its checks. It enforces rules related to data lineage (a), specification (b), external subject fetching (c), and state changes (d), as well as file existence for edits (e).

Makoto is installable as a plugin for Claude Code version 2.1.275 or later via the command /plugin install makoto --marketplace Clear-Sights/Makoto. It requires Python 3.11 or later and relies solely on the standard library. All session state is maintained in a locked, hash-linked journal located at ~/.claude/makoto2_state (configurable via MAKOTO_STATE_DIR). The plugin also supports an optional host envelope for recording observed data, such as opaque mutations or program executions.

Why it matters for builders

For AI builders developing agent-based systems, Makoto offers a robust mechanism for enforcing action integrity. It moves beyond simple warnings to actively block potentially erroneous or deceptive agent actions. By ensuring that agents operate based on verifiable, up-to-date information, Makoto helps to build more trustworthy and predictable AI systems. This is particularly important for agents involved in code generation, modification, or deployment, where factual accuracy and adherence to process are paramount.

Practical impact

Builders can integrate Makoto into their Claude Code workflows to add a layer of deterministic verification. To start using Makoto, ensure Claude Code version 2.1.275 or later is installed, then run the plugin installation command. Developers should familiarize themselves with the specific immediacy rules (a-e) to understand how Makoto enforces data integrity. For agents performing tasks like code commits or file edits, Makoto will automatically check for prerequisite actions like reading the latest version of a file or verifying that changes have been tested. If a violation occurs, the agent's step will be blocked, and Makoto will indicate the specific rule and missing information, providing a clear path to resolution.

Caveats and source limits

The source material indicates that while Makoto's checks are deterministic and aim for zero false positives, the overall effect of its dispatcher on a live session's outcome is unmeasured. The license is FSL-1.1-ALv2, permitting use for most purposes except offering it as a commercial product with similar functionality; it transitions to Apache-2.0 two years after release. Specific pricing for the plugin or any associated services is not provided. The source also does not detail independent benchmarks comparing Makoto's performance or overhead against other verification methods.

Sources

Written with AI assistance from the linked sources; every claim below was checked against them automatically. How we produce articles.

Claim check: 7/7 supported claims - 7 evidence links - 100% avg confidence
  • Makoto is a Claude Code hook that checks agent actions against logged deeds before execution.supported - github.com
  • Makoto enforces principles of 'Immediacy': 'Original' (read original instead of derivation) and 'Unchanged' (do not assume what can change has not).supported - github.com
  • Makoto blocks steps that violate immediacy rules, providing specific reasons and clear resolution paths.supported - github.com
  • Makoto is installable for Claude Code 2.1.275 or later via `/plugin install makoto --marketplace Clear-Sights/Makoto`.supported - github.com
  • Makoto requires Python 3.11 or later and uses only the standard library.supported - github.com
  • Makoto's checker uses no model and no network, aiming for measured zero false positives.supported - github.com
  • The license for Makoto is FSL-1.1-ALv2, transitioning to Apache-2.0 two years after release.supported - github.com

Caveats

  • Single-source caution: verify critical details at the linked source.
Radar score 87/100 - how it was calculated
Reliability82
Freshness92
Novelty77
Technical85
Developer96
Ecosystem72
Confidence96
  • Reliability 82: GitHub metadata supports source trust
  • Freshness 92: Fresh GitHub activity
  • Novelty 77: Fresh GitHub release
  • Technical 85: Repository technical metadata
  • Developer 96: Developer tooling signals
  • Ecosystem 72: Fresh GitHub release
  • Confidence 96: Claims have reliable evidence
Share
XLinkedInHacker News

Discussion

Loading comments...

Related articles

AI Coding - Nov 9, 2026Beluga IDE: AI-Native Mac App with Peer-to-Peer Agent CommunicationBeluga is a new AI-native IDE for macOS designed for AI development, featuring peer-to-peer agent communication and a unique worktree system for managing multiple tasks simultaneously. The latest release, v2.3.2-mac, enhances agent collaboration and task visualization.Agents - Nov 11, 2026Dodaios v4.2.3: Open-Source Superagent for DesktopDodaios, formerly Personal Jarvis, has released version 4.2.3. This open-source superagent runs on your desktop, managing multiple specialized agents for tasks like coding and research. It supports various AI models and offers a gamified interface for managing agent workflows.Agents - Nov 11, 2026Myco Connects AI Agents Across MachinesMyco is a new tool designed to connect various AI agents, including Claude Code and Codex, across different machines. It allows agents to discover each other, delegate tasks, and communicate through a central hub, simplifying multi-agent workflows.Agents - Nov 12, 2026Bernstein: Open-Source AI Agent Governance Layer ReleasedBernstein, an open-source governance layer for AI agents, has been released. It allows developers to define rules declaratively, which Bernstein then enforces to produce verifiable and replayable records of agent actions. The system aims to provide deterministic scheduling and robust auditing capabilities.AI Coding - Nov 9, 2026Claude Code Templates Registry and CLI UpdatedThe davila7/claude-code-templates repository offers an open-source registry and CLI for Claude Code components, including agents, commands, and skills. It recently released version v1.29.7, indicating ongoing development and support for Anthropic's Claude code ecosystem.AI Coding - Nov 11, 2026Claude Code Skills for Chinese DevelopersThis repository offers a curated collection of Claude Code Skills, Agents, and Plugins specifically for Chinese developers. It categorizes these resources by use case, allowing for easy integration and continuous updates.