How OMS works

From one correction to guidance your agents can use

A correction becomes a proposed rule. OMS checks it against existing guidance, applies your review policy and publishes approved changes for the intended teams.

The whole process

Follow one lesson through OMS

From a correction in one task to a skill another agent can use. The graph keeps the rule, its context and its history connected.

One connected process. A reusable lesson.

Contribute, sanitise, compile, publish, distribute, use. The graph holds the history; the installed skill carries the guidance into the next task.

01 / Contribute

People, agents & integrations

Example correction

Before contacting <EMAIL_1> for a quote, check the approved supplier register.

MCP · SDKs · signals · skill imports

02 / Sanitise

Then store

03 / Compile

Distil. Classify. Reconcile.

Read correctionWrite rule

OMS

Shared graph

SourceConstraintSkillRule

One source of truth · linked history

Explore & review

People, where needed.

Inspect the graph. Resolve flagged items under your review policy.

Conditional branch

04 / Publish

Conformance & conflict gates

purchasing / SKILL.md

Check supplier approval before requesting a quote.

+ CLAUDE.md instructions

05 / Distribute

Git pull · CLI / file sync

06 / Use

Your agents. Their next task.

Read the installed skill natively.

No live OMS call to read the file

File delivery shown. Connected agents can also retrieve skills through MCP; that route needs a live connection to OMS. Explore both routes

Contribute
Send a correction or lesson.
Sanitise
Replace detected personal details.
Compile
Reconcile guidance in the graph.
Publish
Check the release; build skill files.
Distribute
Sync files to enrolled users.
Use
Agents read the installed skills.

Illustrative process, not a product interface. Example data.

What becomes guidance

A correction is something a person tells an agent to do differently: use this supplier, never paste that data, run the check before you send. An agent can also propose what it worked out during a task. Either way, the contribution is a proposal, not permission to publish.

Capture

Keep the lesson and its source: OMS records the proposed guidance with the context needed to review it later: which agent, which task, when and who submitted it.
Screen before ordinary storage: Contributions are screened for supported personal-data patterns and prompt injection before storage. Screening does not guarantee detection of every sensitive detail.

Check and approve

Compile: OMS turns the correction into specific, testable guidance. If one correction contains two rules, they can be split so each can be reviewed, scoped and retired separately.
Reconcile: OMS checks whether a proposal adds new guidance, repeats an existing rule or conflicts with it. It can strengthen a matching rule or hold a conflict for a decision instead of creating another copy.
Apply review policy: New and revised guidance follows the configured approval policy. Detected conflicts with locked constraints block publication and must be reviewed before release.

Publish and deliver

Build the skill: Approved rules are published in a reusable skill: a package of guidance that can also contain supporting files.
Deliver to configured teams: Published skills reach enrolled users through configured file updates or a live MCP connection. Team and access settings determine who can receive them.
Keep the history: Each release can be compared with the previous version. Replaced guidance is retired from what ships, and an earlier release can be restored.

Delivery

How do agents receive skills from OMS?

Through skill files

Updated for your users. Loaded when needed.

Once set up, OMS keeps published skills updated for the users and teams connected to it. Compatible agents see which skills are available and load the relevant detail when a task needs it. They read local files, so OMS adds no live service call to skill loading.

Automatic updates: Approved changes reach enrolled users through the configured refresh process.
Only the relevant detail: Compatible agents discover the available skills first and load the contents when a task calls for them. This is progressive disclosure: the long instructions are read on demand, not held in every prompt.
Outside the runtime path: Agents can read their installed skills while OMS is unavailable.

Through a connection

The same knowledge, without native skill support.

Agent platforms that do not read skill files can still use your organisation’s guidance. Through a supported connection called MCP, a configured agent can find and request the skills it needs. The same connection can send useful corrections back to OMS.

Extend access: Bring managed guidance to connected platforms without a native skills reader.
Keep one source: Serve the same managed knowledge through files or a connection.

Refresh depends on a successful installation, the right permissions, connectivity and the configured schedule, and it does not change a session that is already running. Reading installed skills needs no connection; sending a correction or receiving an update does. Live retrieval over MCP needs OMS to be available and a platform that supports the connection.

Technical details

Illustrative file output

These examples show the shape of a published skill. They are not installation instructions. Use the maintained documentation for supported setup steps.

Illustrative configuration — not installation instructions

File names, locations and frontmatter depend on the supported delivery setup. Confirm the current format in Docs before implementation.

Illustrative file layouttext

~/                         # your home directory
  .claude/
    CLAUDE.md              # imports OMS instructions
    skills/
      data-handling/       # linked OMS skill
        SKILL.md
      travel-booking/      # linked OMS skill
        SKILL.md

Default Claude Code layout, under your home directory. OMS adds an instruction import and links published skills here, not in your repository. A custom CLAUDE_CONFIG_DIR changes this location. Use the maintained documentation for setup instructions.

Read the maintained setup documentation

Evaluate the result separately

Our benchmark programme tests instruction-following and the maintenance process. This is a separate evaluation activity, not continuous monitoring of every agent session. It can inform a correction or retirement decision, but OMS does not automatically detect every external change or guarantee that a skill improves outcomes.

See a correction become a rule

See how your team’s corrections would become reviewed, shared guidance.

Book an OMS demo