5. Run it with your agent

The simplest way to run a pipeline is with the coding agent you already use, on your own machine. Kodel hands out one step at a time and records the outcome; your agent does the work in your checkout, under your subscription. No model key inside Kodel, no build infrastructure.

Connect your agent once

Settings → Profile → Access tokens → Create token. Give it a name and tick MCP server (mcp:*). The secret is shown once — copy it.

The same page has a Connect a tool card with ready-to-paste commands for Claude Code, OpenCode, Codex, Cursor, kodelctl and curl, already filled in with your instance address and the token you just created. For Claude Code:

claude mcp add --transport http kodel http://localhost:8080/api/mcp \
  --header "Authorization: Bearer <your-token>"

Then install the driver skill — ask your agent to call the pipeline_init tool and save the file it returns to .claude/skills/kodel-pipeline/SKILL.md. That skill teaches the agent the loop below.

Run the task

In your repository checkout, tell the agent:

Run the pipeline for task DEV-1.

It does the rest:

1. pipeline_resolve — which pipeline applies to this task, and can it run here? For our template the answer is feature-delivery, environment local, runnable_here: true.

2. pipeline_start — creates the run and returns a run_id.

3. pipeline_next_step — asks the server what to do now. The server answers with one step:

   {
     "slug": "plan",
     "kind": "agent",
     "name": "Plan",
     "prompt": "Plan the work: break the task into concrete steps before writing any code.",
     "git": { "run_branch": "kodel/wip/DEV-1", "base_branch": "main" }
   }

4. The agent does that one step, then pipeline_complete_step with the outcome and a short summary. The server advances the graph and the loop repeats — implement, then the test check with its command — until:

   { "done": true, "run_status": "completed" }

Note what the agent is not doing: deciding the order. The server computes the next step from the published graph, so the process is the same whichever agent executes it — and a rerun after a failure resumes rather than starting over.

Branches

The step payload carries the run branch (kodel/wip/DEV-1) and its base. Work there, and when a step reports back it can register the branch it actually used so the task’s Changes tab knows what to diff.

Approvals

When the run reaches an approval step, the agent stops: it prints who is holding the run and what the decision is about, then ends its turn. Approve in the UI (the task page or My Work), then ask the agent to continue — it resumes the same run, it does not start a new one.

Waiting is deliberately opt-in: an agent parked on a gate would burn your tokens polling. If you want it to block and continue by itself, enable auto-wait on the pipeline or the specific step.

Starting a run from the UI

The task page also has a Run ”…” button. That path asks the platform to execute the steps itself, which needs infrastructure to run them on — a dev box or a self-hosted runner. Use it once you have that; for a laptop walkthrough, drive the run from your agent.

Next

Follow the run in the task.