Problem
For cost tracking and post-run analysis of agentic workflows, the missing piece is read access to workflow run artifacts. Specifically, downloading the agent-artifacts artifact from a completed gh aw run to parse token-usage.jsonl for per-run token cost data.
The current workaround requires chaining several REST API calls:
GET /repos/{owner}/{repo}/actions/runs/{run_id}/artifacts to list artifactsGET /repos/{owner}/{repo}/actions/artifacts/{artifact_id}/zip to download- Unzip and parse the relevant file
This is a lot of plumbing for what should be a natural thing for an AI agent to do: "look at what the last run produced."
Proposed tool
download_workflow_run_artifact(owner, repo, run_id, artifact_name) that returns the artifact contents (or a structured representation for known formats like JSONL).
Why this matters
The token-usage.jsonl format written by gh-aw's firewall proxy (since v0.25.8) contains per-model token counts for every agentic workflow run. A single MCP tool call to retrieve it would unlock a natural class of "analyze what the last agent run did" workflows. Right now that analysis requires dropping out of MCP and into raw REST.
More broadly, artifacts are a primary output mechanism for GitHub Actions. Being able to read them via MCP would make a wide range of post-run analysis workflows significantly simpler.
Problem
For cost tracking and post-run analysis of agentic workflows, the missing piece is read access to workflow run artifacts. Specifically, downloading the
agent-artifactsartifact from a completedgh awrun to parsetoken-usage.jsonlfor per-run token cost data.The current workaround requires chaining several REST API calls:
GET /repos/{owner}/{repo}/actions/runs/{run_id}/artifactsto list artifactsGET /repos/{owner}/{repo}/actions/artifacts/{artifact_id}/zipto downloadThis is a lot of plumbing for what should be a natural thing for an AI agent to do: "look at what the last run produced."
Proposed tool
download_workflow_run_artifact(owner, repo, run_id, artifact_name)that returns the artifact contents (or a structured representation for known formats like JSONL).Why this matters
The
token-usage.jsonlformat written by gh-aw's firewall proxy (since v0.25.8) contains per-model token counts for every agentic workflow run. A single MCP tool call to retrieve it would unlock a natural class of "analyze what the last agent run did" workflows. Right now that analysis requires dropping out of MCP and into raw REST.More broadly, artifacts are a primary output mechanism for GitHub Actions. Being able to read them via MCP would make a wide range of post-run analysis workflows significantly simpler.