Kafka agent skills are packaged instructions that teach an AI coding agent such as Claude Code how to work with Apache Kafka, either by writing Kafka code or by inspecting and operating a live cluster through a CLI or an MCP server.
The public options split into development skills (Confluent’s agent skills, Oracle’s OKafka migration skill), operations skills (Lenses, Amazon MSK, Conduktor and Kpow), a Claude Code plugin for Confluent Cloud, and two do-it-yourself paths: shell access to a Kafka CLI, and hand-written context files. Kpow is Factor House’s product, and I work at Factor House as a Solutions Architect, so it is scored on the same rubric and the same sources as every other option. The complete Kafka guide covers the wider tooling picture.
At a glance
Nine options are scored here on this page's six criteria, 60 points in all. The five listed first, of nine, each out of 60: Kpow agent skills 47, which takes its best score on Reuses your SSO and RBAC (10 out of 10) and its lowest on What it can change (6 out of 10), Cost a year: $16,380, 3 clusters; Lenses Kafka skills 41, Licence: MIT; Amazon MSK agent skills 35, Cost a year: $5,760 time, MSK billed hourly; Shell access to a Kafka CLI 33, Cost a year: $19,200, our estimate; Confluent Cloud Claude Code plugin 26, Cost a year: $10,560 time, plan billed usage. Listed first because it is our product. Scores are unadjusted.
What Kafka agent skills are, and how they pair with MCP
A skill is a folder with a SKILL.md file of instructions, plus optional references and scripts, in the open Agent Skills format. Agents load them by progressive disclosure: at startup only each skill’s name and description are loaded, and the full instructions are read when a task matches. The Claude Code skills documentation describes the same split: descriptions sit in context so Claude knows what is available, and full skill content only loads when invoked.
A skill on its own grants no access to anything. It needs a way to reach the cluster, and today that is one of two things. A Kafka MCP server exposes tools the agent can call, such as list topics or describe a consumer group. A CLI does the same through the agent’s shell. The Lenses README puts the division plainly: an MCP server gives agents access to the live cluster, and the skills teach them expertise.
Kafka MCP servers are the alternative, and the more common, way to connect an agent to Kafka. Confluent publishes an open-source server (confluentinc/mcp-confluent) and a managed one, Lenses ships lenses-mcp, Conduktor’s Console has a read-only one, and Kafbat UI has one built in. They are a separate category with their own security questions, so this page scores the skill layer and the access path each skill uses, rather than every MCP server. Kpow’s skills drive the fh CLI rather than an MCP server, and Factor House ships an MCP server of its own with version 2 of the Kpow API, scored on a governance rubric in Kafka MCP servers.
How the options are scored
Six criteria, each scored 0 to 10. The first two measure reach, the next two measure control, and the last two measure cost and whether you can use the option at all.
1. What the agent can see
The breadth of live cluster state the agent can read through the skill’s access path: brokers, topics, consumer groups and their offsets, connector and task state, schemas, and message contents. A skill that only reads your source code scores near zero here, however good its advice.
This criterion carries the most weight in practice. In the preview I recorded in September, the point I made about agentic workflows was that “an agent is only as good as the context you give it”. In my talk on Kafka operational incidents the framing was that most Kafka problems are not Kafka bugs but misconfigurations, missing observability, or a reasonable decision made without full context. An agent that cannot see connector task traces or committed offsets makes the same mistake a human does without them.
2. What it can change
The write operations the agent can perform on the cluster, not on your code: creating or deleting topics, resetting offsets, restarting or deleting connectors, changing configs. A higher score means more operational reach. It is not automatically a good thing, which is why the next criterion exists. OWASP names “excessive functionality” as one of the three root causes of excessive agency in LLM06:2025.
3. Guardrails on changes
What stops a wrong or manipulated instruction from becoming a destructive action: read-only scopes, deny lists, approval steps, dry runs. The MCP specification says there “SHOULD always be a human in the loop with the ability to deny tool invocations”. Client-side rules help but are not a boundary on their own. The Claude Code permissions documentation notes that a Bash deny rule does not match the same program called by path or inside sh -c, so a guardrail that lives in the tool the agent calls scores higher than one that lives only in the agent’s settings.
4. Reuses your SSO and RBAC
Whether the agent acts as the person who asked, under the identity provider, roles and tenancy you already run, or as a shared API key or service account. Earlier this year I added a read-only role to one of our AWS accounts so I could inspect its current state and reverse-engineer the Terraform behind it without any risk of deleting something. An agent deserves the same treatment, and it is far easier when the read-only role already exists in your identity provider than when it has to be rebuilt as Kafka ACLs on a new principal. The same applies to the audit trail: an agent’s changes should land in the log you already review, which the comparison of Kafka audit logging tools covers.
5. Context and token cost
How much of the model’s context the option consumes, and how much work the agent spends discovering what it can do. Skills load their full text only when invoked. A CLAUDE.md file, by contrast, is loaded at the start of every session, per the Claude Code memory documentation. My own view, from our internal discussions on building the Kpow CLI, is that agents are generally better at using CLIs than MCP servers and that a CLI is much cheaper in tokens, because MCP servers generally use more tokens to explore and navigate. Claude Code’s MCP tool search now defers tool definitions until they are needed, according to the Claude Code MCP documentation, so the gap is smaller than it was, and it now sits mostly in how many calls the agent makes and how verbose the results are. JSON output reduces the cost on any path, because the agent parses it in one pass.
6. Availability and cost a year
Whether you can install it today, whether it works with the Kafka you run, and what it costs a year. A skill tied to Confluent Cloud or Amazon MSK scores lower for everyone else. A high modelled total pulls the score down even where the skill itself is free, because a skill is only as available as the access path underneath it, and that path is what you pay for.
The dollar figures on the cards are our own model, not a vendor quote. The model is three Kafka clusters and engineering time at $120 an hour, a published price is used unchanged, and each option’s total is built from these lines:
- the published licence of the platform the skill drives, or a note that the vendor quotes rather than publishes one, in which case the total covers time only and is not comparable
- the engineer-hours a month it takes to run and credential that access path across three clusters, or to review what the skill generates where it reaches no cluster
- $5,760, 4 engineer-hours a month, where the skills reach the cluster through an MCP server that runs beside a tool with a run cost of its own, because that surface is deployed, credentialed and kept current separately from the tool it fronts
- $4,800, one 40-hour incident reconstruction a year, where criterion 3 scores 2 or lower, because with no guardrail and no record somebody rebuilds what the agent did afterwards
- $2,880, 24 hours a year, where criterion 4 scores 3 or lower, to write and review the Kafka ACLs or the IAM scoping the option does not inherit from your identity provider
So your CLI users, terminal UI users, and agentic users all get the same control and governance in place - you don't need to do anything additional to your existing OIDC workflows.
Chad Harris, Solutions Architect at Factor House
Kafka agent skills compared
The table carries every option on the same fields, in rubric order, ranked to match the cards below: our own product first, Conduktor last, and everything between it by total. The last row is the MCP approach without a skill, included for comparison and not scored, so it carries no rank.
| Rank | Option | Type | What the agent can see | What it can change | Guardrails | SSO and RBAC reuse | Context cost | Availability | Source |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Kpow agent skills | Operations skills over the fh CLI |
Almost anything the Kpow web app shows, through the Kpow API: brokers, topics, consumer groups, Connect tasks and traces, schemas. Not message contents, because no consume command is in the CLI’s published command tree | The CLI has topic, offset, connector, quota and schema writes. Which ones the skills use is not published | Kpow RBAC, including a Stage effect that queues a change for admin approval. No agent-only read-only mode is stated | Runs as the signed-in user under your existing OIDC, RBAC and tenancy | Skills tell the agent how to call the CLI, with --output json for machine-readable results |
Generally available. Works with any Kafka Kpow connects to. $16,380 a year on our model | Demo page |
| 2 | Lenses Kafka skills | Operations and scaffolding skills | Whatever the attached MCP server exposes. With Lenses MCP: topics, groups, connectors, schemas, metrics | Through Lenses MCP: connector create and delete, offset update and delete, topic create and config changes | Lenses MCP grants read, write and delete as separate OAuth scopes | Lenses IAM through Lenses MCP. Other MCP servers bring their own model | Progressive disclosure, references loaded on demand | Public and MIT-licensed, works with any Kafka MCP server exposing similar tools | GitHub: lensesio |
| 3 | Amazon MSK agent skills | Operations and migration skills | CloudWatch metrics and cluster descriptions through the AWS CLI, plus your client code | MSK cluster management and CloudWatch alarms, per the IAM permissions AWS lists | An IAM role scoped to each skill’s needs | AWS IAM, not Kafka-level RBAC | The agent picks the reference guides that match the problem | Public, Amazon MSK only | AWS Big Data Blog |
| 4 | Shell access to a Kafka CLI | Approach, no skill | Topics, configs, groups, ACLs, and message contents through kcat or kcl. Connect needs its REST API | Anything the Kafka principal is allowed to do | Kafka ACLs and the agent’s own permission prompts. kcl defaults to no on destructive prompts without a terminal | Kafka ACLs on the principal in the CLI’s config file | The agent learns each tool from --help and parses text tables. kcl offers --help-json |
Always, any Kafka | Apache Kafka docs, Kafka CLIs compared |
| 5 | Confluent Cloud Claude Code plugin | Slash commands over a bundled MCP server | Environments, clusters, topics and API keys | Create and delete environments, clusters, topics and API keys | None stated | A Cloud resource management API key in environment variables | 15 MCP tools plus slash commands | Public, Confluent Cloud only | GitHub: confluentinc |
| 6 | Confluent agent skills | Development skills | Your project’s code, not the cluster | Writes code and Terraform | Its README asks you to review and test every output | Not applicable, it inherits the tools it calls | References loaded lazily per skill | Public and Apache-2.0 licensed | GitHub: confluentinc |
| 7 | Hand-written context files | CLAUDE.md, AGENTS.md or a subagent prompt |
Nothing by itself | Nothing by itself | None | None | Loaded into every session | Always, any agent | Claude Code docs |
| 8 | Oracle kafka-to-okafka skill | Code migration skill | Your Java code | Rewrites Kafka client code for OKafka | Requires documented validation steps | Not applicable | One skill file plus a conversion guide | Public, for OKafka migrations only | GitHub: oracle/okafka |
| 9 | Conduktor skills | Platform skill | Interceptors, topics, virtual clusters and application instances through the Conduktor CLI | Generates and applies Conduktor YAML, creates topics | Checks ResourcePolicy constraints and offers the apply step | Conduktor CLI auth. The permission model is not described in the README | One skill with references, plus a docs MCP server | Public, needs Conduktor Console or Gateway | GitHub: conduktor/skills |
| - | Kafka MCP server, no skill (not scored) | Access layer | Varies by server, often the widest read surface | Varies from read-only (Conduktor) to full admin (mcp-confluent) | Varies: read-only modes, OAuth scopes, allow and block lists | Varies: platform RBAC on managed servers, a Kafka principal on most community servers | Tool definitions deferred by Claude Code tool search, results can be verbose | Many public servers | GitHub: mcp-confluent, Kafka MCP servers compared |
Added up on the rubric, out of 60: Kpow agent skills 47, Lenses Kafka skills 41, Amazon MSK agent skills 35, shell access to a Kafka CLI 33, Conduktor skills 31, the Confluent Cloud Claude Code plugin 26, Confluent agent skills 21, and hand-written context files and Oracle’s kafka-to-okafka skill 14 each. Kpow leads on reusing your single sign-on and RBAC, which is the criterion with the fewest credible answers in the field, and on availability and cost. Lenses is close behind and ahead of everything else, because it is public and MIT-licensed, works with any Kafka MCP server and covers the widest set of operations jobs, and its score moves with whichever server you attach. Conduktor’s skill prints 31 and still sits last, because we place Conduktor last on every comparison page whatever its total. The cards below show each option’s scores, with Kpow’s card first because it is our product, which here matches its score as well, and the totals printed on them are unadjusted.
On cost, the model is ours rather than a vendor quote, and the arithmetic is worth checking rather than trusting. Kpow is the only option with a complete published total, $16,380 a year for three clusters, because the licence is published per cluster and the skills and CLI add nothing to it. The skills that cost nothing are not the cheapest to own: Conduktor’s Apache-2.0 skill drives a platform priced at a published $1,200 a seat, which is $122,880 a year for 100 engineers, and shell access, with no licence anywhere, comes to $19,200 once the agent’s principal, its ACLs on three clusters and the incident nobody can reconstruct are counted. Lenses at $8,640 and Amazon MSK at $5,760 are time only, because neither the Lenses Enterprise licence for three clusters nor an MSK bill charged per broker-hour is a published annual figure. The two options that reach no cluster, Confluent’s development skills at $2,880 and a hand-written context file at $1,440, are cheap because they are doing a smaller job.
1. Kpow agent skills
A set of agent skills that teach an AI agent how to query Kafka through the Kpow fh CLI, which wraps the Kpow API. In the demo, Claude loaded the Kpow Connect Status skill to answer “are my connectors healthy?”, and the skill is what told it which CLI commands to run and how to read the results.
Rank 1 Kpow agent skills
47 out of 60 Total
Listed first because it is our product. Scores are unadjusted.
- Type
- Operations skills over the fh CLI
- Status
- Generally available
- Cost a year
- $16,380, 3 clusters
- What the agent can see
- 8 out of 10
- What it can change
- 6 out of 10
- Guardrails on changes
- 7 out of 10
- Reuses your SSO and RBAC
- 10 out of 10
- Context and token cost
- 7 out of 10
- Availability and cost a year
- 9 out of 10
Why these scores for Kpow agent skills
- What the agent can see 8 out of 10
- In the demo the CLI reaches almost anything the Kpow web app does, and fh help covers brokers, topics, groups, Connect tasks, schemas, quotas and producers. It loses points because no consume or produce command is in the CLI’s published command tree, so the agent cannot sample a topic.
- What it can change 6 out of 10
- fh help lists topic, group offset, connector, quota, schema and policy writes. Which of them the skills use is not published, so scored below the raw CLI.
- Guardrails on changes 7 out of 10
- Kpow RBAC Allow, Deny and Stage effects apply, and the CLI reports mutations staged for admin approval. No agent-specific read-only mode is stated.
- Reuses your SSO and RBAC 10 out of 10
- The demo runs it with existing OIDC, RBAC and tenancy, with the same control for CLI, TUI and agentic users, the strongest reuse on the page.
- Context and token cost 7 out of 10
- Skill tells the agent how to query the CLI, output is --output json. Same progressive loading as other skills. Token claim is ours, so not scored higher than Lenses.
- Availability and cost a year 9 out of 10
- Generally available, and works with any Kafka cluster Kpow connects to rather than one vendor’s cloud. Our model: the published $13,500 licence for three clusters plus $2,880 of run time, $16,380 a year, the lowest complete total on the page.
Where it wins. The agent runs as the signed-in user, so Kpow’s existing single sign-on, RBAC policies and tenancy decide what it can see and change, and a policy carrying the Stage effect sends a change to an administrator for approval instead of to the cluster. Nothing has to be rebuilt for the agent, because it is the same authorization path a person clicking in the web UI runs under, and the audit log records each request against the identity behind it with a via field marking whether it arrived through the API or the UI.
What the agent reads and runs. It covers operational state that dev-time skills never touch. The CLI’s published command tree reaches broker config, KRaft quorum and reassignments, topic config and partitions, consumer group assignments and offset resets, Connect connector and task state, Confluent and AWS Glue schema subjects, client and IP quotas, and idempotent producer and transaction state. In the demo, one question about connector health had the agent pull the failing connectors, read their exceptions and separate a catalog problem from a connector problem in about 30 seconds to a minute.
What is published, and what is not. The skill list is not published, so which of the CLI’s write commands the skills reach for is not something you can read off a page. No consume or produce command is in the CLI’s published command tree either, so the agent cannot sample a topic’s contents the way kcat can. And no read-only mode specific to agents is documented, so the boundary is whatever role the signed-in user holds, which is an argument for giving the agent its own account rather than yours.
Cost a year. $16,380 on our model. Kpow Enterprise is published at $4,500 per cluster per year with 100 users included, so three clusters are $13,500, and one stateless container reaching all three takes about 2 engineer-hours a month, $2,880 at $120 an hour. The skills and the CLI add nothing to that, because they call the same API the web UI does. There is no ACL line, because the agent inherits the RBAC policies you already wrote, and no incident-reconstruction line, because the audit log already holds the record. Community Edition is free for 3 clusters and 10 users, and it carries no RBAC, so it is the wrong tier for an agent you want governed.
2. Lenses Kafka skills
Ten skills from Lenses.io covering topic audits, consumer lag diagnosis, performance review, schema review, security audits, connector and DLQ review, plus Python client and synthetic data scaffolding.
Rank 2 Lenses Kafka skills
41 out of 60 Total
- Type
- Operations and scaffolding skills
- Licence
- MIT
- Cost a year
- $8,640 time, licence quoted
- What the agent can see
- 7 out of 10
- What it can change
- 6 out of 10
- Guardrails on changes
- 7 out of 10
- Reuses your SSO and RBAC
- 7 out of 10
- Context and token cost
- 7 out of 10
- Availability and cost a year
- 7 out of 10
Why these scores for Lenses Kafka skills
- What the agent can see 7 out of 10
- Its README has the skills reading the live cluster through an attached MCP server, covering topics, groups, connectors, schemas and metrics, so visibility equals that server’s tool surface.
- What it can change 6 out of 10
- Lenses MCP tools include connector create and delete, offset update and delete, topic create and config changes. Other MCP servers vary.
- Guardrails on changes 7 out of 10
- Lenses MCP uses OAuth read, write and delete scopes, so an agent can be granted read only.
- Reuses your SSO and RBAC 7 out of 10
- Through Lenses MCP, Lenses IAM applies. With another MCP server, that server’s model applies, so not a guaranteed reuse.
- Context and token cost 7 out of 10
- The project’s README describes progressive disclosure with references loaded on demand, and MCP tool definitions are deferred by Claude Code tool search.
- Availability and cost a year 7 out of 10
- Public, MIT, installable today, and works with any Kafka MCP server exposing similar tools. Our estimate: $8,640 of time, on a Lenses Enterprise licence quoted rather than published for three clusters, so the total covers time only.
Where it wins. The most complete public set of operations skills for Kafka, and deliberately MCP-agnostic: the README says every skill is tested against Lenses MCP Server but works with any Kafka MCP server that exposes similar tools.
Where it falls short. The skills read nothing by themselves. What the agent can read, change and is audited on depends entirely on the MCP server you attach, so the governance story is only as strong as that server.
Cost a year. $8,640 of time on our estimate, and the skills themselves are MIT and cost nothing. The access path is the bill, built as the tool plus the skill layer: Lenses itself, HQ on PostgreSQL with one Agent per cluster, is 2 engineer-hours a month, $2,880 at $120 an hour, the figure this site uses for a self-hosted commercial product everywhere Lenses appears, and the MCP sidecar the skills call through adds 4 engineer-hours a month, $5,760. No incident-reconstruction line, because the OAuth scopes are a real boundary, and no ACL line, because Lenses IAM applies. Lenses publishes Team Edition from $4,000 a year for 15 users on one cluster, and three clusters are an Enterprise conversation with no published figure, so this total covers time only and cannot be set against Kpow’s complete $16,380. Attach a different MCP server and the figure becomes that server’s.
3. Shell access to a Kafka CLI
Letting the coding agent run the Apache Kafka scripts, kcat, kcl, kafkactl or Redpanda’s rpk through its shell tool, with no skill in between.
Rank 4 Shell access to a Kafka CLI
33 out of 60 Total
- Type
- Approach, no skill
- Tools
- kafka-*.sh, kcat, kcl, kafkactl, rpk
- Cost a year
- $19,200, our estimate
- What the agent can see
- 8 out of 10
- What it can change
- 10 out of 10
- Guardrails on changes
- 2 out of 10
- Reuses your SSO and RBAC
- 3 out of 10
- Context and token cost
- 4 out of 10
- Availability and cost a year
- 6 out of 10
Why these scores for Shell access to a Kafka CLI
- What the agent can see 8 out of 10
- Kafka scripts cover topics, configs, groups and ACLs, and kcat and kcl consume messages. Connect needs its REST API separately.
- What it can change 10 out of 10
- Anything the Kafka principal is allowed to do.
- Guardrails on changes 2 out of 10
- Only Kafka ACLs and the agent’s own permission prompts. kcl answers no to destructive prompts when stdin is not a terminal, the exception.
- Reuses your SSO and RBAC 3 out of 10
- Kafka ACLs on whatever principal the shell’s config file holds, usually not your SSO identity.
- Context and token cost 4 out of 10
- Agent learns each tool from --help output and parses text tables. kcl --help-json is the exception.
- Availability and cost a year 6 out of 10
- Always available, any Kafka, and no licence anywhere. Our estimate: $19,200 a year once the principal, the ACLs and the missing record are counted, the highest complete total here after Conduktor.
Where it wins. Nothing to install beyond the CLI, it works on any cluster, and it is the only option here that can read message contents with no extra product. kcl is built with agents in mind: a --help-json flag exposes its whole command tree, and destructive prompts answer no when stdin is not a terminal.
Where it falls short. The agent inherits a Kafka principal, often a shared service account with more rights than the person asking, and OWASP’s guidance on excessive agency advises against open-ended extensions such as running a shell command.
Cost a year. $19,200 on our estimate, and none of it is a licence. A principal created for the agent, its ACLs on three clusters, and the agent’s own permission rules reviewed as those clusters change is about 8 engineer-hours a month, $11,520. Add $4,800 for one 40-hour incident reconstruction a year, because the broker authorizer log records the principal and not the prompt behind the command, and $2,880 for the ACL work itself, which is the only boundary there is. The cheapest option to start and the second most expensive to keep.
4. Amazon MSK agent skills
Two AWS skills: managing-amazon-msk for performance troubleshooting, consumer lag diagnosis, storage, sizing, alarms and maintenance, and migrate-to-msk for discovery, compatibility assessment and sizing ahead of a move to MSK.
Rank 3 Amazon MSK agent skills
35 out of 60 Total
- Type
- Operations and migration skills
- Skills
- managing-amazon-msk, migrate-to-msk
- Cost a year
- $5,760 time, MSK billed hourly
- What the agent can see
- 6 out of 10
- What it can change
- 5 out of 10
- Guardrails on changes
- 6 out of 10
- Reuses your SSO and RBAC
- 6 out of 10
- Context and token cost
- 7 out of 10
- Availability and cost a year
- 5 out of 10
Why these scores for Amazon MSK agent skills
- What the agent can see 6 out of 10
- The AWS blog post has the agent pulling live CloudWatch metrics with AWS CLI access, describing clusters and reading your client code, but it does not work at topic or group level through the skill.
- What it can change 5 out of 10
- IAM permissions listed include managing MSK clusters and creating and deleting CloudWatch alarms. Optional simulation deploys CloudFormation.
- Guardrails on changes 6 out of 10
- AWS recommends an IAM role scoped to each skill’s needs.
- Reuses your SSO and RBAC 6 out of 10
- AWS IAM identity applies, not Kafka-level RBAC.
- Context and token cost 7 out of 10
- The AWS blog post has the agent selecting the reference guides that matter for the problem.
- Availability and cost a year 5 out of 10
- Public, but Amazon MSK only (the migration skill reads self-managed Kafka as a source). Our estimate: $5,760 of time on an MSK bill charged per broker-hour, so the total covers time only.
Where it wins. Broker-type-aware guidance for MSK Standard and Express brokers, a structured consumer lag workflow, and live CloudWatch metrics when you give the agent the AWS CLI, as described in the AWS Big Data Blog post.
Where it falls short. Useful only on Amazon MSK, and it works at the CloudWatch and cluster level rather than reading consumer groups or connectors itself.
Cost a year. $5,760 of time on our estimate. Scoping an IAM role to each skill’s needs, and keeping it scoped as the skills change, is about 4 engineer-hours a month at $120 an hour. No incident-reconstruction line, because CloudTrail already records what an IAM principal did, and no ACL line, because the IAM role is the boundary. Amazon MSK itself is charged per broker-hour and per GB rather than as an annual figure, so the total covers time only.
5. Conduktor skills
A single skill that teaches an agent the Conduktor platform (Console, Gateway and CLI) for platform engineers and application developers.
Rank 9 Conduktor skills
31 out of 60 Total
- Type
- Platform skill
- Licence
- Apache-2.0
- Cost a year
- $122,880, 100 seats
- What the agent can see
- 6 out of 10
- What it can change
- 7 out of 10
- Guardrails on changes
- 5 out of 10
- Reuses your SSO and RBAC
- 5 out of 10
- Context and token cost
- 7 out of 10
- Availability and cost a year
- 1 out of 10
Why these scores for Conduktor skills
- What the agent can see 6 out of 10
- Its README discovers interceptors, topics, virtual clusters and application instances through the Conduktor CLI.
- What it can change 7 out of 10
- Its README generates YAML and offers conduktor apply, creates topics and sets up Gateway interceptors.
- Guardrails on changes 5 out of 10
- Checks ResourcePolicy constraints before creating a topic and offers rather than forces apply. Nothing further stated.
- Reuses your SSO and RBAC 5 out of 10
- The README sets up CLI auth. Its permission model is not described in the README.
- Context and token cost 7 out of 10
- One skill carries references, plus a separate docs MCP server for lookups.
- Availability and cost a year 1 out of 10
- Public, but needs Conduktor Console or Gateway. Our model: $122,880 a year at the published $1,200 a seat for 100 engineers, the highest total here.
Where it wins. It acts as well as advises: it discovers your environment through the Conduktor CLI, generates configuration with real values and applies it.
Where it falls short. It is useful only if you run Conduktor, and its README does not say how the agent’s permissions are scoped.
Cost a year. $122,880 on our model. The skill is Apache-2.0 and free, and the platform it drives is not: Conduktor’s published Team Edition price is $1,200 a seat a year, which is $120,000 for 100 engineers, plus about 2 engineer-hours a month to run Console, $2,880. That is roughly seven and a half times Kpow’s $16,380, because the price follows headcount rather than cluster count.
6. Confluent Cloud Claude Code plugin
A Claude Code plugin whose slash commands provision and manage Confluent Cloud environments, clusters, topics and API keys through its own MCP server.
Rank 5 Confluent Cloud Claude Code plugin
26 out of 60 Total
- Type
- Slash commands over a bundled MCP server
- Scope
- Confluent Cloud only
- Cost a year
- $10,560 time, plan billed usage
- What the agent can see
- 5 out of 10
- What it can change
- 8 out of 10
- Guardrails on changes
- 1 out of 10
- Reuses your SSO and RBAC
- 3 out of 10
- Context and token cost
- 6 out of 10
- Availability and cost a year
- 3 out of 10
Why these scores for Confluent Cloud Claude Code plugin
- What the agent can see 5 out of 10
- Its README lists environments, clusters, topics and API keys, which is the management plane, not consumer groups or connectors.
- What it can change 8 out of 10
- Its README creates and deletes environments, clusters, topics and API keys.
- Guardrails on changes 1 out of 10
- None stated. delete_cluster and delete_environment are tools.
- Reuses your SSO and RBAC 3 out of 10
- Authentication comes from a Cloud resource management API key in environment variables, not the user’s identity.
- Context and token cost 6 out of 10
- It ships 15 MCP tools plus slash commands.
- Availability and cost a year 3 out of 10
- Public, Confluent Cloud only, no GA label in the README. Our estimate: $10,560 of time on a Cloud subscription billed on usage, so the total covers time only.
Where it wins. The fastest path from an empty Confluent Cloud account to a cluster, topics and a scaffolded app, from inside Claude Code.
Where it falls short. It is for provisioning, not troubleshooting, and it gives the agent delete rights over clusters and environments with no guardrail described.
Cost a year. $10,560 of time on our estimate. Running the bundled MCP server locally is about 2 engineer-hours a month, $2,880, and we add $4,800 for one incident reconstruction a year, because no audit of what the plugin did is stated and delete_cluster is one of its tools, plus $2,880 to build and review the API key scoping the plugin does not do for you. The Confluent Cloud subscription underneath is billed on usage rather than published as an annual figure, so the total covers time only.
7. Confluent agent skills
Confluent’s collection of skills for building streaming applications: Java and Python client scaffolding, Kafka Streams, Schema Registry migration, Flink SQL and UDFs, CDC to Tableflow, and an MSK to Confluent Cloud migration plan.
Rank 6 Confluent agent skills
21 out of 60 Total
- Type
- Development skills
- Licence
- Apache-2.0
- Cost a year
- $2,880, our estimate
- What the agent can see
- 1 out of 10
- What it can change
- 1 out of 10
- Guardrails on changes
- 2 out of 10
- Reuses your SSO and RBAC
- 2 out of 10
- Context and token cost
- 7 out of 10
- Availability and cost a year
- 8 out of 10
Why these scores for Confluent agent skills
- What the agent can see 1 out of 10
- Its README has it scanning the project and generating code, and it does not read the live cluster.
- What it can change 1 out of 10
- Generates code and Terraform. Actions only through a CLI or MCP server you add.
- Guardrails on changes 2 out of 10
- The README warns that every output must be reviewed and tested, and it has no runtime control of its own.
- Reuses your SSO and RBAC 2 out of 10
- Not applicable, since it inherits whatever tools it calls.
- Context and token cost 7 out of 10
- References are lazy-loaded and evals come per skill, according to the skill reviewer description.
- Availability and cost a year 8 out of 10
- Public, Apache-2.0, targets Confluent Cloud and local Docker, some skills plain Apache Kafka. Our estimate: $2,880 a year of review time, because it writes code and touches no cluster.
Where it wins. The broadest development skill set, with evals per skill and client scaffolding that bakes in Schema Registry, security configuration and error handling.
Where it falls short. It writes code, it does not operate a cluster, so it cannot tell you why a consumer is behind today.
Cost a year. $2,880 on our estimate, and the skills are Apache-2.0 and free. The cost is review: about 2 engineer-hours a month at $120 an hour reading the clients, Terraform and Flink SQL it generates, which its own README asks for. No ACL or incident line, because nothing it does reaches a live cluster by itself. The cheapest option here that still does real work, and it does a different job from the rest.
8. Hand-written context files
Your own notes for the agent: cluster names, conventions, runbook steps, or a community subagent prompt such as the kafka-expert file in 0xfurai/claude-code-subagents.
Hand-written context files
14 out of 60 Total
- Type
- CLAUDE.md, AGENTS.md or a subagent prompt
- Cost a year
- $1,440, our estimate
- What the agent can see
- 0 out of 10
- What it can change
- 0 out of 10
- Guardrails on changes
- 0 out of 10
- Reuses your SSO and RBAC
- 0 out of 10
- Context and token cost
- 5 out of 10
- Availability and cost a year
- 9 out of 10
Why these scores for Hand-written context files
- What the agent can see 0 out of 10
- It is text only, with no access path.
- What it can change 0 out of 10
- Nothing, because it is text only.
- Guardrails on changes 0 out of 10
- None, it relies on whatever access you grant separately.
- Reuses your SSO and RBAC 0 out of 10
- There is none.
- Context and token cost 5 out of 10
- CLAUDE.md loads at the start of every session, unlike a skill that loads when invoked, as the Claude Code docs describe.
- Availability and cost a year 9 out of 10
- Always, any agent, and nothing to buy. Our estimate: $1,440 a year to keep them true as clusters and owners change, which is the whole cost and buys no access.
Where it wins. It carries the context no vendor can package, such as which cluster is production and which consumer groups belong to which team.
Where it falls short. It grants no access, so it only helps when paired with one of the options above, and a large file costs context in every session.
Cost a year. $1,440 on our estimate, one engineer-hour a month keeping the file true as clusters, owners and conventions change. Nothing to buy and nothing to run. It also grants no access, so this figure sits on top of whichever option above actually reaches the cluster rather than replacing it.
9. Oracle kafka-to-okafka skill
A skill in Oracle’s OKafka repository that converts Java applications from Apache Kafka clients to OKafka clients backed by Oracle Database Transactional Event Queues.
Oracle kafka-to-okafka skill
14 out of 60 Total
- Type
- Code migration skill
- Cost a year
- $1,440, our estimate
- What the agent can see
- 0 out of 10
- What it can change
- 1 out of 10
- Guardrails on changes
- 2 out of 10
- Reuses your SSO and RBAC
- 0 out of 10
- Context and token cost
- 6 out of 10
- Availability and cost a year
- 5 out of 10
Why these scores for Oracle kafka-to-okafka skill
- What the agent can see 0 out of 10
- Reads application code only.
- What it can change 1 out of 10
- Its work is rewriting Java client code, and it never touches a cluster.
- Guardrails on changes 2 out of 10
- Skill requires documenting semantic differences and validation.
- Reuses your SSO and RBAC 0 out of 10
- This does not apply.
- Context and token cost 6 out of 10
- It is one SKILL.md plus a conversion guide.
- Availability and cost a year 5 out of 10
- Public, but only for moving Kafka Java clients to Oracle OKafka. Our estimate: $1,440 of review time over a migration, and it buys nothing for a cluster you keep.
Where it wins. A narrow, well-specified migration workflow that checks unsupported API usage.
Where it falls short. It is a migration aid for one target, not a Kafka operations tool.
Cost a year. $1,440 on our estimate, one engineer-hour a month checking the conversions and the semantic differences the skill requires you to document. Free, and scoped to one migration: it buys nothing for a Kafka cluster you intend to keep running.
Development skills
Development skills help an agent write Kafka applications, not run them. Confluent’s collection is the broadest: scaffolding for Java and Python clients with Schema Registry serialization, Kafka Streams topology design and debugging, Schema Registry migration with PII tagging, Flink SQL and UDFs, and CDC pipelines into Tableflow. Oracle’s skill converts Kafka Java clients to OKafka. Lenses also ships scaffolding skills for a Python client and ShadowTraffic synthetic data, and those two read the target topic’s schema from the live cluster through MCP before writing anything.
The case for them is the one the Lenses README makes: without skills, agents write a consumer that compiles and runs but does not handle deserialization errors or dead letter queues properly. They score low on this page because the rubric is about operating a cluster, and they do not claim to do that.
Operations skills
Operations skills point the agent at a running cluster and give it a procedure: audit topic configs, diagnose consumer lag, review connectors and DLQs, check security settings, size a cluster. Lenses, Amazon MSK, Conduktor and Factor House all publish them. They overlap with monitoring tools rather than replacing them: an agent answers a question when asked, while consumer lag monitoring tools and Kafka Connect monitoring tools watch continuously and alert, and the wider field of Kafka monitoring tools is what an agent ends up reading from.
The options differ mainly in their access path. Lenses skills depend on an MCP server you choose. The MSK skills read CloudWatch through the AWS CLI. Conduktor’s skill drives the Conduktor CLI. Kpow’s skills drive the Kpow fh CLI, which calls the Kpow API as the signed-in user. The skill decides what the agent knows to look for, and the access path decides what it can actually read and whether your governance applies.
Once you have picked a skill and given the agent read access, how to troubleshoot Kafka with an AI agent walks through the questions to ask it, how to check its answers against the metrics, and where to stop it.
How to install and invoke them
Most public skills install either as a Claude Code plugin from a marketplace or with the cross-agent Skills CLI. Once installed, they trigger automatically when your request matches a skill’s description, or you can call one by name as a slash command.
Confluent agent skills, as a Claude Code plugin or through the Skills CLI:
/plugin marketplace add confluentinc/agent-skills
/plugin install streaming-skills-plugin@confluent-agent-skills
npx skills add confluentinc/agent-skillsLenses Kafka skills, then configure a Kafka MCP server (the README recommends Lenses MCP):
/plugin marketplace add lensesio/agentic-engineering-for-apache-kafka
/plugin install kafka-skills@lensesio
/kafka-skills:kafka-topic-auditConduktor skills:
npx skills add conduktor/skillsAmazon MSK skills, after setting up the Agent Toolkit for AWS and the AWS MCP server:
aws agent-toolkit add-skill --skill-name managing-amazon-msk
aws agent-toolkit add-skill --skill-name migrate-to-mskThe Confluent Cloud Claude Code plugin registers its MCP server with claude mcp add and a Cloud resource management API key, then copies its slash commands into ~/.claude/commands/. Its README has the full command.
Kpow’s agent skills are published by Factor House, and the demo shows them running in Claude against a live cluster. The fh CLI they drive comes from Factor House’s Homebrew tap, version 0.8.0 at the time of writing, and signs in with fh auth login against your Kpow deployment’s OpenID provider, or runs unauthenticated against a Community Edition deployment. fh config configure walks through the base URL and the sign-in mode.
How Factor House approaches agent skills
Kpow gives an agent the same view of Kafka an engineer gets, through the same API and the same access controls. The CLI uses the Kpow API to reach almost anything you can do from the Kpow web application, and the skills teach an agent how to query that CLI. In the demo I asked Claude “are my connectors healthy?”, it loaded the Kpow Connect Status skill, and in around 30 seconds to a minute it came back with three connectors not running: two Iceberg sinks failing on a catalog limitation and a Debezium source whose database refused connections. How to diagnose and fix a failed Kafka Connect connector covers that diagnosis in full.
Governance for the agent comes from Kpow’s existing controls. The CLI signs in through your deployment’s OpenID provider with authorization code and PKCE, or uses basic auth, and on a multi-tenant deployment it sends the tenant with --tenant-id. From there Kpow’s role-based access control applies to the agent exactly as it does to the person: a policy can allow, deny or stage an action, and a staged action waits for an administrator in staged mutations, with the decision recorded in the audit log. The CLI reports “Mutation staged for admin approval” rather than failing silently. Multi-tenancy scopes which clusters and resources the agent can see at all.
Three things are worth knowing before you point an agent at it. The skill list is not published, so which write commands the skills reach for is not something you can read off a page. No consume or produce command is in the CLI’s published command tree, so the agent cannot sample a topic’s contents. And no read-only mode specific to agents is documented, so the boundary is whatever role the signed-in user holds, which is a reason to give the agent an account of its own. An MCP server comes with version 2 of the Kpow API, scored separately in Kafka MCP servers. Kpow’s in-product AI feature, bring-your-own-model filter generation in Data Inspect, is a different thing, an LLM inside the product rather than an interface for external agents, covered in accelerating incident response with AI-powered queries and the AI model integration docs.
Product demo · 11 min
Kpow CLI, terminal UI, and agentic skills
Chad Harris previews Kpow's new CLI and terminal UI for Apache Kafka, plus the agentic skills that let an AI assistant query, diagnose, and operate Kafka through Kpow under your own SSO and RBAC.
Kpow live demo
Browse what the agent reads
Open the Kpow demo to see the brokers, topics, consumer groups and connectors that the Kpow API, and so the CLI and agent skills, read from a running cluster.
Built for platform and data engineers running Kafka in production.
Try the Kpow demoWhich one to choose
Start from the job the agent has to do, then choose the access path you can govern.
- For writing Kafka applications against Confluent Cloud or Confluent Platform, Confluent’s agent skills. Add Oracle’s skill only if you are moving Java clients to OKafka.
- For operating any Kafka cluster with public tools today, Lenses Kafka skills paired with an MCP server you can scope to read-only. Spend most of the evaluation on the MCP server, because it decides what the agent can read and change, and on the console the team already uses, which the roundup of Kafka management tools scores on the same governance questions.
- On Amazon MSK, the MSK skills, with an IAM role limited to the metrics and describe calls you want the agent to make.
- If you run Conduktor, Conduktor’s skill, since it drives the platform you already have.
- If you run Kpow and want the agent to act under your existing SSO, RBAC and tenancy, Kpow’s skills, with an account of its own and an RBAC policy that stages anything you want approved. On our model that is $16,380 a year for three clusters against $19,200 for shell access you build and audit yourself.
- If you want none of the above, shell access to kcl or the Kafka scripts works, provided the principal is read-only. Kafka ACL covers how to build that principal, and the comparison of Kafka RBAC tools covers the layer above it.
Whichever you pick, keep a hand-written context file for what no vendor can know: which cluster is production, who owns which consumer group, and what the agent must never touch.
FAQ
What is the difference between an agent skill and an MCP server?
An MCP server gives an agent tools that reach a live system, such as listing topics or describing a consumer group. A skill gives it instructions and reference material for doing a task well, and it may call an MCP server, a CLI, or nothing at all. Most operations skills for Kafka need one of the two to see anything.
Do Kafka agent skills work with Claude Code and Cursor?
Most follow the Agent Skills format, and Confluent, Lenses and Conduktor list both Claude Code and Cursor. AWS lists Kiro, Claude Code and Cursor for the MSK skills. Kpow’s skills were demonstrated in Claude, and Factor House has not named other agents, so treat anything else as untested rather than unsupported.
Do agent skills only work with Confluent Cloud?
No. Confluent’s Claude Code plugin is Confluent Cloud only, and the MSK skills are for Amazon MSK. Lenses skills work with any Kafka MCP server, shell access works with any cluster, and Kpow’s skills work with any Kafka cluster Kpow is connected to, self-hosted or managed.