// AI agent release radar · updated live
The live release feed for AI agents.
Track new MCP servers, agent SDKs, model releases, APIs, tools, and libraries. Every update explains what agents can do now and links back to the source.
// live MCP demo
Run the real anonymous list_recent_capabilities tool.
No account required to browse · Read-only MCP works without a key · JSON · RSS · Atom
// latest AI agent capabilities
New capabilities, live
live wire
all events →- libraryDSPy 3.4.1: structured GEPA InstructionProposer default and RLM sub-agentsstanfordnlp — An agent can now use GEPA's structured InstructionProposer by default (preserving images/history/tool trajectories as structured content) and have dspy.RLM spawn DSPy sub-agents (Predict, ChainOfThought, ReActV2, nested RLM) from inside its sandboxed REPL under a shared max_llm_c8h ago
- mcpMCP reference servers hit v1.0.0 (filesystem, memory, fetch, git, time, sequential-thinking)modelcontextprotocol — An agent can now install the stable 1.0.0 reference MCP servers (filesystem, memory, sequential-thinking, everything, fetch, git, time) with hardened session handling, roots refresh, and JSON Schema 2020-12 tool schemas.2d ago
- sdkAnthropic Python SDK v1.13.0 adds Managed Agents workflows & multiagent configanthropic — An agent can now configure multiagent setups, define workflows, and filter threads by status via the Managed Agents API, plus read Chat/Cowork unified analytics metrics through the Python SDK.3d ago
- cliClaude Code v2.1.295: blocking hooks, OSC 7501 status, gateway model allowlistsAnthropic — An agent can now block actions when hooks fail via `onFailure: "block"`, use Program Status Protocol (OSC 7501) for terminal state signaling, restrict upstreams to specific models, and cap retry wait times via `CLAUDE_CODE_RETRY_WATCHDOG_MAX_WAIT_MS`.4d ago
- cliClaude Code v2.1.293 adds Claude Haiku 5.5 with 1M context as default Haikuanthropics — An agent can now use Claude Haiku 5.5 with a 1M-token context via `claude-haiku-5-5`, register mod tools with `isDeferred:false` to surface their schema upfront, and read `agentType` from subagent status lines.5d ago
- sdkAnthropic Python SDK v1.12.0 adds claude-haiku-5-5 and typed computer/browser tool callsAnthropic — An agent can now call claude-haiku-5-5 and use typed computer/browser tool calls, configure url_sources on the Managed Agents web_fetch tool, and filter models by lifecycle stage via the official Python SDK.5d ago
- modelAnthropic releases Claude Haiku 5.5Anthropic — An agent can now call Claude Haiku 5.5, Anthropic's fastest and cheapest small model, for high-volume cost-sensitive tasks via the Anthropic API.6d ago
- cliClaude Code v2.1.292 adds Agent effort control, plugin marketplace install flag, and prompt caching for modsAnthropic — An agent can now dispatch sub-agents at a chosen effort level via the Agent tool's `effort` parameter, install plugins directly from a marketplace source in one command, and cache prompt/system blocks when mods call `$.model.complete`.6d ago
- sdkOpenAI Python SDK 3.25.0 adds agent tool-failure observability and streamed local tool runsopenai — An agent can now observe local tool failures, run local tools during streamed session creation, group usage by source, and use deferred tool-search loading via typed function helpers in the OpenAI Python SDK.6d ago
- libraryHugging Face Transformers v5.19.0 adds EmbeddingGemma 2 multimodal embeddings & expert-parallel dispatchhuggingface — An agent can now produce unified multimodal embeddings across text, images, audio, and video with EmbeddingGemma 2 via Transformers, with truncatable Matryoshka dims for cheaper cross-modal retrieval and clustering.6d ago
// pipeline
How the feed works
01
Discover
We scan GitHub releases, package registries, official MCP listings, arXiv, HN, first-party vendor news, and publisher-verified skill metadata on a recurring schedule.
02
Classify
The curator extracts what changed, what an agent can do now, package identity, and a minimal example.
03
Broadcast
One canonical CapabilityEvent, available through MCP, REST, JSON Feed, RSS, and Atom.
// pricing
Plans for humans and agents.
Buy a subscription on behalf of your agent. Machine-callable checkout is a first-class path — an agent can subscribe by JSON POST.
Free
$0/ mo
Enough to prove your agent will actually read the wire.
- ▸Public web + RSS / Atom / JSON
- ▸REST API + MCP
- ▸100 event API + MCP calls / day
- ▸1 named API key
- ▸Per-agent delivery tracking
most agents pick this
Agent Pro
$19/ mo
A serious allowance for an always-on agent.
- ▸MCP server (streamable HTTP)
- ▸REST API + JSON feed
- ▸RSS + Atom feeds
- ▸10,000 event API + MCP calls / day
- ▸Up to 5 named API keys
- ▸Per-agent delivery tracking
Agent Team
$99/ mo
Higher limits and more keys for an agent fleet.
- ▸Everything in Agent Pro
- ▸100,000 event API + MCP calls / day
- ▸Up to 25 named API keys
- ▸Fleet-scale MCP + REST access
- ▸Per-agent delivery tracking
// docs preview
Three ways in.
Point your agent at MCP for tool-shaped access, or hit the REST API from any language.
MCP (Claude Desktop / Cursor)
~/.config/mcp/config.json
{
"mcpServers": {
"agentwire": {
"transport": "streamable_http",
"url": "https://api-production-22727.up.railway.app/mcp"
}
}
}curl
shell
curl -s https://api-production-22727.up.railway.app/v1/events?limit=5 \
-H "Authorization: Bearer $AGENTWIRE_KEY"Python (stdlib, zero deps)
pull_wire.py
import os, urllib.request, json
req = urllib.request.Request(
"https://api-production-22727.up.railway.app/v1/events?limit=5",
headers={"Authorization": f"Bearer {os.environ['AGENTWIRE_KEY']}"},
)
events = json.loads(urllib.request.urlopen(req).read())["data"]
for e in events:
print(e["kind"], "-", e["title"])Base URL: https://api-production-22727.up.railway.app