Agent Dynamo authoring plugins

Coding-harness plugins for the Agent Dynamo platform. Currently one plugin: agent-authoring — author, run, and debug agents, workflow agents, and work queues directly from your harness, with zero local install.

Supported harnesses: Claude Code and Codex. One marketplace, one plugin, one shared skill; each harness reads its own manifest.

Install

Claude Code

/plugin marketplace add darugar/agent-dynamo-authoring-plugins
/plugin install agent-authoring@agent-dynamo

Codex

codex plugin marketplace add darugar/agent-dynamo-authoring-plugins
codex plugin add agent-authoring@agent-dynamo

Then set your API key in the environment the harness runs in (mint one in the Agent Dynamo web app under Settings → API keys):

export AGENTDYNAMO_API_KEY=ad_...

Add that line to your shell profile (~/.zshrc / ~/.bashrc) so it survives new terminals, and restart the harness. That's it — the plugin bundles the connection to the hosted agent-dynamo MCP server, so its tools appear automatically, along with an agent-authoring skill that guides spec writing.

What you get

• MCP connection to https://app.agentdynamo.com/mcp, authenticated with your key — every tool call is scoped to your account. Tools include building_blocks, validate_spec, apply_spec, run_agent, execution_log, enqueue_items, queue_status, queue_control, export_agent, and authoring_guide (the full authoring reference, fetched on demand — no docs to install).
• Skill (agent-authoring) that tells the model how to author AgentSpecs: consult the guide first, look up real building-block names, then validate → dry-run → apply → run → inspect.

Try: "Create an agent that summarizes a URL I give it, then run it on example.com".

Without the plugin

Any MCP client can connect directly — same endpoint, same key:

# Claude Code
claude mcp add --transport http agent-dynamo https://app.agentdynamo.com/mcp \
  --header "Authorization: Bearer ad_..."

# Codex
codex mcp add agent-dynamo --url https://app.agentdynamo.com/mcp \
  --bearer-token-env-var AGENTDYNAMO_API_KEY

Other clients (Cursor, claude.ai custom connectors) take the same URL — /mcp exactly, no trailing slash — with an Authorization: Bearer ad_... header.

Headless & CI use

Claude Code. Interactive sessions prompt for tool permissions the first time — nothing to configure. But non-interactive runs (claude -p, the Agent SDK, CI) and shared permission allowlists in .claude/settings.json must name the tools explicitly, and plugin-bundled MCP servers get a namespaced permission scope (plugin_<plugin>_<server>), not the server name alone:

claude -p "List my Agent Dynamo agents" \
  --allowedTools "mcp__plugin_agent-authoring_agent-dynamo__*"

Two gotchas: a bare mcp__* wildcard is rejected (allow rules must name a server scope; globs are only valid in the tool position), and allowlists written for a manually added server (mcp__agent-dynamo__* via claude mcp add) do not match the plugin's tools. Hook matchers follow the same scoped naming.

Codex. codex exec prompts for MCP tool approval and cancels the call if nothing answers — approval_policy = "never" does not auto-approve it. Expect user cancelled MCP tool call in unattended runs unless you explicitly relax approvals for the session.

Troubleshooting

• "Missing Authorization header" — AGENTDYNAMO_API_KEY isn't set in the environment the harness started from. Set it and restart the harness.
• "Authentication failed (403)" — the key is wrong, revoked, or belongs to a different instance (a key minted against a local dev instance will 403 against app.agentdynamo.com). Mint a new one under Settings → API keys.

Layout

.claude-plugin/marketplace.json          Claude Code marketplace catalog
.agents/plugins/marketplace.json         Codex marketplace catalog
plugins/agent-authoring/
  .claude-plugin/plugin.json             Claude Code manifest
  .codex-plugin/plugin.json              Codex manifest
  .mcp.json                              MCP server, Claude Code shape
  .mcp.codex.json                        MCP server, Codex shape
  skills/agent-authoring/SKILL.md        shared by both harnesses

The two MCP files are not interchangeable: Claude Code uses {"type": "http", "headers": {"Authorization": "Bearer ${VAR}"}}, while Codex uses {"url": ..., "bearer_token_env_var": "VAR"}. The Codex manifest points at its own file via "mcpServers": "./.mcp.codex.json".

${VAR} interpolation is Claude-only. Claude Code expands ${VAR} and ${VAR:-default} in .mcp.json at connect time, so its URL is ${AGENTDYNAMO_MCP_URL:-https://app.agentdynamo.com/mcp} — set AGENTDYNAMO_MCP_URL to point the plugin at a different instance.

Codex does not expand ${VAR} in plugin MCP config; it passes the string through literally, and a URL containing ${...} fails to connect at all. So .mcp.codex.json must hard-code the URL — do not copy the Claude pattern into it. A Codex user who needs a different endpoint registers the server by hand instead:

codex mcp add agent-dynamo --url https://your-instance.example.com/mcp \
  --bearer-token-env-var AGENTDYNAMO_API_KEY

(Verified by pointing each harness at a local HTTP server and observing which URL it actually requested — codex mcp get and claude mcp get both display the uninterpolated string, so they cannot be used to test this.)

Maintainers

This repo owns the skill. There is no copy in the app repo and nothing to sync. That works because the skill deliberately holds only stable procedure — "call authoring_guide first, look up real building-block names, then validate → dry-run → apply → run → inspect" — while every version-specific detail reaches the model at runtime through the MCP tools the platform serves: authoring_guide (the full authoring reference), building_blocks (real model ids and tool names), spec_schema (live JSON Schema), and validate_spec's error hints (including migration hints for removed spec fields).

So a platform change usually needs no release here. Edit this skill only when the authoring procedure changes — a new step in the loop, a renamed tool, a changed connection story.

To release: bump version in both plugins/agent-authoring/.claude-plugin/plugin.json and plugins/agent-authoring/.codex-plugin/plugin.json — users only see updates on a version bump — then commit and push. Validate before pushing:

claude plugin validate .
codex plugin marketplace add .   # then: codex plugin add agent-authoring@agent-dynamo