CLAWPOD ENGINEERING

Using the Clawpod Capability Registry from a Room

A Clawpod Capability is a pre-verified, reusable combination of instructions and controlled execution tools for reliable agent work.

ClawPod Engineering · 4min read
REVIEWED BYEditorial Review · Capability Review
A person’s request passing through an agent to the right capability at a workbench

What is a Clawpod Capability?

A Clawpod Capability is a pre-verified, reusable combination of instructions and controlled execution tools that helps an agent perform a particular kind of work reliably. Instead of inventing a method every time, the agent selects a unit whose intended requests and supported scope are already documented.

A Capability can include either or both of these elements:

  • A Skill gives the agent instructions for judging the situation and following the right procedure.
  • A CLI Harness lets the agent operate software consistently within a controlled scope.

Users do not need to manage these parts directly. Describe the result you want in a Room, and the agent can select the appropriate Capability.

What does the Capability Registry do?

The Capability Registry is the canonical catalog an agent checks when it needs to find, install, or update a Capability. The agent first checks whether an installed Capability can handle the request. If not, it searches the Registry for a candidate whose scope fits.

flowchart TD
  A[Request the outcome in a Room] --> B[Check installed Capabilities]
  B --> C{Suitable Capability available?}
  C -->|Yes| F[Perform work with the Capability]
  C -->|No| D[Find candidates in the Registry]
  D --> E[Explain choice, then install and validate]
  E --> F
  F --> G[Report result and evidence]

A Registry search is useful when:

  • the agent says no suitable Capability is installed;
  • a repeatable specialist procedure is needed for research, publishing, or an external service;
  • several similar Capabilities need to be compared;
  • you want to check for a verified alternative before creating something new; or
  • you need a newer version or a recoverable earlier version.

Start with this prompt:

First check whether a Capability for this work is already installed. If not, search the canonical Registry and explain the candidates and your recommendation.

Ask the agent to find and compare

You do not need to know a product or internal package name. State the outcome and important constraints.

Find a Capability that can cross-check multiple sources and produce citable research. Check installed Capabilities first. If none fits, compare the three best Registry candidates.

The answer should include whether each candidate is installed or newly found, its exact version, intended requests, supported and unsupported work, selection rationale, and any account or permission requirements. The goal is not the first search hit; it is the candidate with the best matching scope and boundary.

Ask for a pre-install review

Before installing this Capability, check compatibility with the current Clawpod environment, linked Capabilities, required permissions and services, the version to be installed, and the validation method.

A linked Skill and Harness may ship together. The agent should not call the work complete when only part of that linked Capability is ready.

Ask for installation and validation

Install the recommended Capability in the managed environment and validate its integrity and connections. Run a limited-impact check, then report the installed version and evidence.

A useful completion report states:

  • the Capability and exact installed version;
  • whether all linked components are ready;
  • whether the installation matches the Registry declaration;
  • whether the agent can now select it;
  • the result of a limited-impact check; and
  • any remaining account connection or operating condition.

“Installed” and “ready for real work” are not always the same. The agent should make that distinction explicit.

Ask for the outcome, not the package name

Once the Capability is ready, request the result directly:

Cross-check the three main claims in this document against official sources. Separate any conflicting evidence and report citable support with confidence.

If you want visibility into routing, add:

First tell me in one line which Capability you selected and why.

Ask for a safe update

Check whether a newer version exists. Compare scope, compatibility, permissions, and risk with the installed version. Preserve a recovery point, then update and revalidate.

The report should identify old and new versions, changed selection scope or requirements, revalidation results, and rollback availability. Keep the last known-good version until validation succeeds.

Ask for rollback when validation fails

The Capability updated just now is not working correctly. Roll back to the preserved last known-good version, then revalidate linked components and selection readiness. Report the cause and recovery result.

The agent should report the restored version, any linked components restored with it, revalidation results, and preserved changes. It must not silently overwrite user modifications.

When the agent must stop and ask

Finding, installing, and validating a Capability is separate from using it for a risky action. The agent should explain the impact and confirm authorization before:

  • using a password, API key, or other secret;
  • accessing a person’s account or private data;
  • sending a message, publishing, or making a payment;
  • changing production;
  • deleting or overwriting data; or
  • granting new permissions or widening access.

A publishing Capability can be correctly installed while publication itself still requires separate approval. If the target or authorized scope is unclear, the agent should stop before the external action.

What a good completion report answers

  • What was selected, and why did it fit this request?
  • Which exact version is ready?
  • What was validated, and what passed?
  • How was it used for the requested outcome?
  • What limitations or approvals remain?
  • Which known-good state is available for recovery?

The user’s job is not to type internal commands. State the desired result and safety conditions in the Room, then ask the agent to report its selection reasoning and evidence for search, comparison, installation, validation, use, update, or rollback.

Sources