Prompts & resources

Source: orchestration/mcp-gateway/prompts.ts (handlePromptList, handlePromptGet), orchestration/mcp-gateway/resources.ts (handleResourceList, handleResourceRead), orchestration/mcp-gateway/server.ts (resource templates, handler wiring), orchestration/mcp-gateway/skill-dist.generated.ts, orchestration/mcp-gateway/llms-routes.ts, MCP SDK 1.29.0 (types.js McpError, shared/protocol.js).

Besides its 22 tools, the server offers one prompt and five resources plus three resource templates. They are read-only and run nothing, and they work on https://mcp.plungeai.com/v1 and on the same server at https://api.plungeai.com/v1/mcp. All content is live: registry listings come from the registry at request time, and workflow and execution resources are read from your account with an ownership check. There is no static manifest.

12.1 The /plungeai prompt

There is exactly one prompt. It activates the whole platform in one step. Earlier recipe prompts (research pipeline, market research, …) were removed on purpose: they showed up as separate slash commands and locked the client into one behaviour, when the tools already cover everything (prompts.ts).

Using it in a client. Clients list MCP prompts in their slash-command or prompt menu. In Claude Code the prompt appears as /plungeai:plungeai (MCP): type /plungeai to filter the menu and pick it, or run /mcp__plungeai__plungeai directly (the middle part is the server name you chose in claude mcp add). Claude Code splits any text after the command on whitespace, one word per argument, so only the first word would reach request. Run the prompt bare, then type your request as the next message. Other clients may offer a form field for request instead.

• run bare (no request): the model is told to ask you what you want to accomplish;
• with request set (for example research Tesla's Q2, from a client that offers the field): the model is told to handle that request now.

The prompt is a convenience, not a requirement. You can always just talk to your client normally.

prompts/list:

curl -s https://mcp.plungeai.com/v1 \
  -H "Authorization: Bearer $PLUNGE_API_KEY" \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"prompts/list"}'

prompts/get returns description: "PlungeAI MCP" and one user message (type text). It runs nothing, so it is free:

curl -s https://mcp.plungeai.com/v1 \
  -H "Authorization: Bearer $PLUNGE_API_KEY" \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":2,"method":"prompts/get","params":{"name":"plungeai","arguments":{"request":"research Tesla'"'"'s Q2"}}}'

What the message tells the model:

• It now has the full platform through the plungeai_* tools: about 90 active agents, CNL workflows, missions, long-term memory and execution history.
• Discover. "Show me my agents/workflows" means saved workflows (plungeai_list_workflows). The registry (plungeai_list_agents with a natural-language search) is for building blocks. Fetch a full card with agent_id before first use.
• Execute. plungeai_execute_agent (one agent, optional session_id), plungeai_execute_workflow (saved workflow_id or ad-hoc CNL YAML), plungeai_run_mission (bounded autonomous loop with memory).
• Results. plungeai_get_result, plungeai_get_workflow_status (poll async runs), plungeai_executions, and plungeai_followup / plungeai_continue for finished or paused runs.
• Approvals. Relay any "⏸ AWAITING USER APPROVAL" or "⏸ AWAITING USER" block verbatim, never approve on its own, then call plungeai_continue with approve: true or the user's words.
• Build. plungeai_build_workflow generates and saves a workflow. plungeai_workflow is direct CRUD. plungeai_memory is long-term memory.
• Always pass user_request, the user's words verbatim.
• The last line is either Now handle this request: <request> (the argument is trimmed) or Ask the user what they want to accomplish. when request is empty or absent.

An unknown prompt name is rejected with -32602. The SDK prefixes the message with MCP error -32602::

curl -s https://mcp.plungeai.com/v1 \
  -H "Authorization: Bearer $PLUNGE_API_KEY" \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":3,"method":"prompts/get","params":{"name":"research"}}'

12.2 Resources at a glance

resources/list returns five fixed resources (handleResourceList):

resources/templates/list returns three URI templates (server.ts) for resources addressed by a parameter:

The resources and templates carry no size or annotations. The templates carry no description. Every resources/read answers { "contents": [ { "uri", "mimeType", "text" } ] } with one entry. The server's resources capability is {}: no subscribe and no listChanged (Protocol & transport).

curl -s https://mcp.plungeai.com/v1 \
  -H "Authorization: Bearer $PLUNGE_API_KEY" \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":4,"method":"resources/list"}'

curl -s https://mcp.plungeai.com/v1 \
  -H "Authorization: Bearer $PLUNGE_API_KEY" \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":5,"method":"resources/templates/list"}'

Resources or tools? A resource is the same data some tools return, in a form your client can attach to a conversation without a tool call. For searching and filtering, use the tools: plungeai_list_agents for a semantic search, plungeai_list_workflows for folders and search, plungeai_get_result for one task's output (Discovery, Runs, results & follow-ups).

12.3 plungeai://agents/list

The live registry of active agents: kind=agents&status=active&format=md&limit=100. The Markdown starts # Registry search, then *<n> results*. Each agent follows under a ## Name (id) heading with its kind, category and status line, its description and a Good at list. On 2026-09-24 the registry listed 92 active agents (live-verified). The listing is large, about 150 KB, so attach it only when you really want the whole catalog. For a normal question, a filtered plungeai_list_agents search is far smaller.

curl -s https://mcp.plungeai.com/v1 \
  -H "Authorization: Bearer $PLUNGE_API_KEY" \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":6,"method":"resources/read","params":{"uri":"plungeai://agents/list"}}' \
  | cut -c1-1200

12.4 plungeai://agents/categories

Active agents (up to 500) grouped by category into one table, largest first. An agent with no category counts as general:

# Agent Categories (<n> agents, live registry)

| Category | Agents |
|:---|---:|
| google-workspace | … |
…

_Read `plungeai://agents/{category}` for the agents in one category._

This is the cheap way to answer "what kinds of agents are there". On 2026-09-24 the registry held 92 active agents in 14 categories (live-verified): google-workspace, tools, search, microsoft-365, ai, blockchain, communication, ecommerce-agents, payment, entertainment, media, financials, shopping, coding-agents. Note the finance category is financials, not finance.

curl -s https://mcp.plungeai.com/v1 \
  -H "Authorization: Bearer $PLUNGE_API_KEY" \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":7,"method":"resources/read","params":{"uri":"plungeai://agents/categories"}}'

12.5 plungeai://personas/list

The persona registry: kind=personas&format=md&limit=100, as registry Markdown (a ## Name (id) heading per persona, its kind and category, Good at). Unlike the agent resources, it has no status:active filter, because no persona card carries that tag and the filter would empty the list (resources.ts). The list stops at 100: on 2026-09-24 the registry answered *100 results*. For a specific persona, search with plungeai_list_agents {"kind":"personas","search":"…"}. A persona id goes on plungeai_execute_agent persona or a mission's persona (Agents & missions).

curl -s https://mcp.plungeai.com/v1 \
  -H "Authorization: Bearer $PLUNGE_API_KEY" \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":8,"method":"resources/read","params":{"uri":"plungeai://personas/list"}}' \
  | cut -c1-800

12.6 plungeai://workflows/list

Your workflows: the 50 most recently updated saved workflows on your account (what users call "my agents"). One bullet each (bold name, the id as code, then the description) under # Your Workflows (<n>). With no saved workflows, the text is _No workflows yet._. It carries no folders and no search. For those, use plungeai_list_workflows.

curl -s https://mcp.plungeai.com/v1 \
  -H "Authorization: Bearer $PLUNGE_API_KEY" \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":9,"method":"resources/read","params":{"uri":"plungeai://workflows/list"}}'

12.7 plungeai://docs/workflow-authoring

The complete CNL workflow-authoring skill as one Markdown document: how to author, validate and save CNL workflows, with examples. It is a static bundle built into the server and needs no data access. The content is identical to GET https://mcp.plungeai.com/llms-full.txt, which needs no key (61,309 bytes on 2026-09-24, live-verified). Attach it when you want your client to write CNL YAML by hand (Workflows). The same skill is in the catalog at https://skills.plungeai.com.

curl -s https://mcp.plungeai.com/v1 \
  -H "Authorization: Bearer $PLUNGE_API_KEY" \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":10,"method":"resources/read","params":{"uri":"plungeai://docs/workflow-authoring"}}' \
  | cut -c1-600

12.8 plungeai://agents/{category}

The active agents in one category, up to 100, as registry Markdown under # Registry search — category=<category>. The category is passed to the registry as-is, so an unknown or misspelled category is not an error. It returns an empty listing (*0 results*). Take category names from plungeai://agents/categories.

curl -s https://mcp.plungeai.com/v1 \
  -H "Authorization: Bearer $PLUNGE_API_KEY" \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":11,"method":"resources/read","params":{"uri":"plungeai://agents/financials"}}'

12.9 plungeai://workflows/{id}

The saved CNL YAML of one of your workflows, as text/yaml. The id is the workflow id from plungeai_list_workflows or plungeai://workflows/list. The ownership check comes first: an id that does not exist and an id owned by someone else get the same -32002 answer, so you cannot probe for other accounts' ids. A workflow whose YAML body is missing from storage returns the text No YAML content found.

# needs one of your workflow ids — not auto-verified
curl -s https://mcp.plungeai.com/v1 \
  -H "Authorization: Bearer $PLUNGE_API_KEY" \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":12,"method":"resources/read","params":{"uri":"plungeai://workflows/<workflow id>"}}'

12.10 plungeai://executions/{id}

The stored result of one of your runs, by execution id (not the workflow id, Troubleshooting). It reads the same stores as plungeai_get_result: the stored result, then the final task's output. The document looks like this:

# Execution <execution id>

- **Workflow:** `<workflow id>`

## Result

<the run's final output>

When there is no output, the Result section says why:

This resource gives the final result only. For one task's output, the full conversation thread, the "Steps in this run" index or a paused run's ⏸ block, use plungeai_get_result (Runs, results & follow-ups). An unknown id, or one that is not yours, gets -32002 Execution not found: <id>.

# needs one of your execution ids — not auto-verified
curl -s https://mcp.plungeai.com/v1 \
  -H "Authorization: Bearer $PLUNGE_API_KEY" \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":13,"method":"resources/read","params":{"uri":"plungeai://executions/<execution id>"}}'

12.11 Resource errors

resources/read answers JSON-RPC errors, not outcomes. The SDK prefixes McpError messages with MCP error <code>: (types.js).

curl -s https://mcp.plungeai.com/v1 \
  -H "Authorization: Bearer $PLUNGE_API_KEY" \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":14,"method":"resources/read","params":{"uri":"plungeai://nope"}}'

Registry-backed reads time out after 10 seconds. For the full error vocabulary, see Errors, feedback & limits.