Developer path

Try K² coding context in 10 minutes

Load the public Java R&D bundle, connect the K² MCP server to your coding agent, and verify that the agent receives scoped, cited evidence from docs, source, tests, and guide corpora.

Quickstart steps

Commands are intentionally explicit so a developer can reproduce the public bundle path.

Clone and inspect the public bundle

The bundle contains sanitized JSONL corpora for docs, code, and guides, plus MCP, agent, feed, and pipeline examples.

git clone https://github.com/knowledge2-ai/k2-coding-context-demo.git

Run the offline checks

Verify the harness, sample scorer, and configuration without contacting K2 or any external model.

PYTHONPATH=src python -m unittest discover -s tests
PYTHONPATH=src python -m k2_java_rd_demo.cli show-config
PYTHONPATH=src python -m k2_java_rd_demo.cli score-sample

Load the demo corpus into K2

Use your own K2 project and API key. Keep credentials in environment variables or a local ignored .env file. If you do not have a project yet, create one in the K2 Console.

export K2_API_KEY="<your-k2-api-key>"
export K2_API_HOST="https://api.knowledge2.ai"
PYTHONPATH=src python -m k2_java_rd_demo.cli bootstrap-customer-demo --execute

Connect your coding agent over MCP

Use the same stdio server entry for Codex, Claude Code, Cursor, or any MCP-capable agent. Inline examples are below; the full config is in the asset bundle.

docs/customer-demos/demo-java-customer/k2-assets/examples/mcp-config.example.json

Run one cited Java R&D query

Ask for a Flink REST-handler implementation plan and verify the answer cites guide, docs, source, and test evidence.

Which guide rules, source files, and tests should I use before adding includeMissing to JobVertexWatermarksHandler?

MCP config snippets

These examples all point at the same K² stdio MCP server. Use the format your coding agent expects and keep keys in environment variables.

{
  "mcpServers": {
    "k2-java-rd": {
      "command": "python",
      "args": ["scripts/k2_java_rd_mcp_server.py"],
      "env": {
        "K2_API_KEY": "${K2_API_KEY}",
        "K2_API_HOST": "https://api.knowledge2.ai",
        "K2_PROJECT_ID": "${K2_PROJECT_ID}"
      }
    }
  }
}

{
  "mcpServers": {
    "k2-java-rd": {
      "command": "python",
      "args": ["scripts/k2_java_rd_mcp_server.py"],
      "env": {
        "K2_API_KEY": "${K2_API_KEY}",
        "K2_API_HOST": "https://api.knowledge2.ai",
        "K2_PROJECT_ID": "${K2_PROJECT_ID}"
      }
    }
  }
}

[mcp_servers.k2-java-rd]
command = "python"
args = ["scripts/k2_java_rd_mcp_server.py"]
env_vars = [
  "K2_API_KEY",
  "K2_API_HOST",
  "K2_PROJECT_ID",
  "K2_FLINK_DOCS_CORPUS_ID",
  "K2_FLINK_CODE_CORPUS_ID",
  "K2_GUIDES_CORPUS_ID"
]

Full config Cursor config Claude Code config

Expected response

The first cited query should return a concise plan with role-separated evidence, not a generic Java web answer.

Successful shape

Use the REST handler guide first: add a MessageQueryParameter, keep default behavior unchanged, and extend the focused handler test.
Implementation anchors: JobVertexWatermarksHandler, JobVertexWatermarksHeaders, JobVertexWatermarksInfo, and WebMonitorEndpoint registration in flink-runtime.
Test anchor: JobVertexWatermarksHandlerTest should cover default, partial, and include-missing behavior.
Citations should include one guide rule, one Flink 2.2 REST-doc URL, the handler source file, and the neighboring test file.

What to watch for

Answers that suggest Spring MVC, servlet, or JAX-RS controllers are off the project guardrails.
Answers without handler source and neighboring test citations are not using the intended context.
Answers that change default behavior fail the backwards-compatibility requirement.

Next steps

Once the demo query works, the useful question is how the same pattern maps to your own repo.

Point the same loader at one small internal repository and one guide page tree.

Read the pilot report to see how to freeze tasks, guide checks, and scoring before indexing.

Use the pipeline spec as the reviewable contract for what your coding agent can retrieve.

If something does not work

Most quickstart failures are configuration or indexing state, not model behavior.

API key rejected

Confirm the key belongs to the selected K2 project and is exported in the shell that starts the MCP server.

MCP server not visible

Restart the coding agent after editing the MCP config and check that the stdio command runs from the repo root.

No cited results

Wait for ingestion and indexing to finish, then verify corpus IDs in the environment.

Quota or rate-limit errors

Reduce the demo query loop or move the project to the tier used for the pilot.

Credential handling

Do not commit API keys, project IDs from private tenants, or raw evaluator dumps. Use environment variables locally, rotate keys after demos, and keep live scorecards outside public bundles unless they are scrubbed.

Open Knowledge²