π€ Connect your agent¶
Bring Agent Engineering HQ into the workspace you already use. Start with local MCP, then use the same requests over hosted MCP, A2A or HTTP when a service is running.
| Your starting point | Next step |
|---|---|
| A coding agent or editor | Set up local MCP |
| A running remote AgentEng service | Connect over HTTP or use the stdio bridge |
| An A2A client or your own integration | Use typed questions and requests |
AgentEng is the Agent Engineering Conference and event guide from Agent Engineering HQ. This service covers London and San Francisco only, including their published historical events. discover highlights the next published event in each city, or the most recent event when there is no upcoming one. Event names, dates, speakers and registration links come from the public website catalogue.
The official hosted origin is live at https://a2a.agentengineering.world. Installing this package does not start that service. Local lookup work offline and require no provider key.
Protocol connection instructions¶
agenteng connect a2a prints the public agent card, JSON-RPC endpoint and complete curl examples. agenteng connect mcp --transport http prints hosted MCP configuration; omit --transport http for local stdio. agenteng connect acp prints optional installation and local coding-agent session instructions. These commands do not claim a connection has been established, launch agents or change configuration.
Local MCP¶
Install the package with its mcp extra using your package manager or a release wheel. Ensure agenteng is on the client's executable path, or substitute the absolute path to your installed executable. For a source checkout:
Print client-specific instructions without changing any client settings:
agenteng connect codex
agenteng connect claude-code
agenteng connect cursor
agenteng connect generic
For an installed executable, the coding-client commands are:
Codex uses shared MCP configuration for its CLI and IDE integration. See the official OpenAI documentation. Claude Code supports local command servers; see its MCP documentation.
Cursor can use this entry in .cursor/mcp.json. A generic mcpServers client can use the command and arguments in its own supported configuration location:
See Cursor's MCP documentation. Configuration formats differ between clients; agenteng connect CLIENT prints the corresponding shape. These instructions do not install a plugin, register a marketplace listing or authenticate anyone.
Hosted MCP and a stdio bridge¶
Use the official host below, or replace the origin with your own self-hosted service:
agenteng connect codex --transport http --url https://a2a.agentengineering.world
agenteng connect claude-code --transport http --url https://a2a.agentengineering.world
agenteng connect cursor --transport http --url https://a2a.agentengineering.world
The HTTP MCP endpoint is /mcp/ and uses Streamable HTTP. For clients that need a local command, the same installed CLI can bridge to a hosted service:
{
"mcpServers": {
"agenteng": {
"command": "agenteng",
"args": [
"--remote",
"https://a2a.agentengineering.world",
"mcp"
]
}
}
}
This bridge requires network access and relays requests to that origin; it does not silently fall back to a local snapshot. The bridge exposes the same public read-only operations; personal storage and submissions are unsupported.
Useful questions and requests¶
- βWhat is the next Agent Engineering event in San Francisco?β
- βShow Agent Engineering HQ events in San Francisco.β
- βWhich London sessions cover evaluation or agent harnesses?β
- βShow the published London agenda.β
Start with {"operation":"discover"} or {"operation":"events","upcoming":true}. Use returned IDs for agenda, speakers, tickets and planning. London 2026 is agenteng-london-2026; the October San Francisco event is sf-code-engineering-2026, whose published title remains Code Engineering: From Coding Agents to Software Factories.
Event-day operations: {"operation":"live","at":"2026-10-16T10:40:00+01:00"} returns a now/next venue-screen snapshot (at is optional and also works for now and next), {"operation":"bingo","seed":7,"size":5,"format":"svg"} deals a reproducible talk bingo card (the card is always in data; format text, svg or html also returns a printable copy in artifact, and json, the default, leaves artifact empty), {"operation":"guide","section":"arriving"} returns one section of the published London day guide (omit section for all of it, or use "section":"list" for the index), {"operation":"about","section":"connect"} returns the organiser, chair and connection details, and {"operation":"hq","section":"manifesto"} returns Agent Engineering HQ content. City filters accept website slugs such as san-francisco.
An LLM client can use the same typed operations through MCP, or translate them to POST /v1/query using /openapi.json. A2A clients discover /.well-known/agent-card.json and send A2A 1.0 JSON-RPC to / with A2A-Version: 1.0. The card advertises skills, examples. Client/model usage may have its own costs; default AgentEng lookup make zero server-side model calls.
Tool directory¶
Use the same single MCP tool for disciplines, tools and tool:
A2A data parts and HTTP bodies use the inner object. Continue with next_offset and unchanged filters. Plain βList memory toolsβ questions route to directory browsing. Names, aliases, categories and links are bundled offline; no model call or third-party installation is needed. See full directory usage and attribution.
Discovery surfaces¶
The hosted service provides crawlable HTML at / and /events/EVENT_ID, individual event JSON-LD, /events.json, /llms.txt, /robots.txt and /sitemap.xml. The tool directory adds /tools, /tools/TOOL_ID and the full /tools.json export. Human-readable pages link to the official registration source; JSON includes catalogue freshness and supporting evidence.
These surfaces make the service accessible to crawlers and agents. They do not guarantee indexing, search ranking, recommendations by an LLM or automatic discovery in a coding client. An llms.txt is a reading guide, not an instruction override or a registry submission. Event markup preserves published facts; see Google's event structured-data guidance.
The website already maintains its own metadata, sitemap and llms.txt. Release staging adds an AgentEng agent guide and an event feed for website publication. Link the guide from the website and its existing llms.txt after the service is deployed; preserve the site's existing event and organizer identities.
Website A2A chat¶
The conference website's native Talk to AgentEng A2A panel discovers the
agent card and sends A2A 1.0 JSON-RPC SendMessage requests. The service permits
the A2A-Version header in browser CORS preflights. Structured catalogue
requests stay model-free. Free-text questions use operation=chat when the
agent advertises the agenteng-chat skill; older agents still receive ask.
Enable OpenRouter in Cloud Run¶
On the agenteng Cloud Run service, edit the revision's variables and secrets:
| Variable | Value |
|---|---|
AGENTENG_ENABLE_CHAT |
1 |
AGENTENG_MODEL_BASE_URL |
https://openrouter.ai/api/v1 |
AGENTENG_MODEL |
openrouter/free or a compatible model ID |
AGENTENG_MODEL_API_KEY |
Secret Manager reference containing the OpenRouter API key |
AGENTENG_ENABLE_STANDARD |
0 |
AGENTENG_ENABLE_RLM |
0 |
AGENTENG_ENABLE_INTAKE |
0 |
Grant the Cloud Run runtime identity access to that specific secret, deploy the
revision and route traffic to it. Public chat does not require or expose an
operator token. AGENTENG_ENABLE_CHAT is separate from the operator-only
standard/RLM engines. Never put a provider key in website VITE_* variables.
The tag deployment script preserves these manually configured variables and
secret references; container defaults keep optional capabilities off.
Chat can make up to two model calls and two read-only catalogue lookups per question. It can read published events, speakers, talks, agendas and engineering tools. It cannot access private records, databases, arbitrary URLs, files or code execution. Source IDs must come from retrieved evidence; this validates attribution membership, not the factual quality of generated prose.
If the key or flag is absent, the model is busy, provider limits/credits are exhausted, a provider call fails, output is invalid, or the request reaches its 20-second deadline, the existing catalogue answer is returned automatically. Provider failures pause model attempts for five minutes per process before trying again; invalid output pauses them for 30 seconds. A process admits up to five model questions per minute and one concurrent model worker. Multi-instance hosting needs shared edge limits for larger traffic.
The website sends at most six recent user/assistant messages, each up to 2,000
characters, for follow-up questions. History is untrusted data, not instructions.
The service does not retain it. Visitor questions and these recent replies are
sent to the configured inference provider when chat is enabled; no private
organizer records are included. Common unpublished contact identifiers,
credentials, explicit self-disclosures and requests for private records are
refused before a model call. This is not a complete personal-data detector.
OpenRouter calls require provider.data_collection=deny and provider.zdr=true;
unavailable eligible endpoints use static answers without relaxing these rules.
Disable gateway prompt logging and training opt-ins in your OpenRouter account
privacy settings. See privacy details. Clear chat starts fresh.
The default allowed browser origin is https://agentengineering.world.
For a local website running at http://localhost:8080, configure the local
service explicitly:
AGENTENG_PUBLIC_URL=http://127.0.0.1:8000 AGENTENG_ALLOWED_ORIGINS=http://localhost:8080 agenteng serve
Set VITE_AGENTENG_A2A_URL=http://127.0.0.1:8000 in the website's .env.local.
Use the actual browser origin if your development port differs.
Public chat supplies brief, sourced protocol introductions rather than searching broad FAQs for basic MCP/A2A/ACP/CLI questions. Greetings, "How are you?" and thanks use no model or catalogue search; a simple βtell me moreβ follow-up uses recent user context. Identity, agent engineering definitions and setup questions also have focused public fallback answers. General search ignores conversational filler and excludes attendance FAQs unless the visitor asks an attendance question. When existing evidence supports an answer, the model receives a final JSON request without tools. Otherwise it may perform one additional public lookup round (at most two lookups), then must answer; irrelevant FAQ calls are rejected. Cards follow cited lookup results, not the last tool call. Provider exceptions never appear in replies; usage.fallback_reason is a bounded code (disabled, unconfigured, busy, provider_limit, provider_unavailable, or invalid_response). Invalid output has a 30-second cooldown; provider limits/outages use five minutes.