The agentic web's git log. A signed, append-only announcement layer at
/.well-known/agent-feed.xml, plus a public observatory of how the existing ecosystem disagrees with itself.
Live site → abdul-abdi.github.io/agent-feed · Drift dashboard · SPEC.md · IETF draft · MCP SEP
Agent endpoints in 2026 are described across MCP server cards, A2A registries, agents.json, llms.txt, and a half-dozen GitHub catalogs. They drift. They disagree. When an API mutates underneath an agent, the agent breaks silently — there is no robots.txt-equivalent for telling agents the world has moved.
This is that announcement layer — plus a public observatory of where the existing standards already disagree about the same agent endpoint.
Live: https://abdul-abdi.github.io/agent-feed/ — homepage, full SPEC viewer, IETF draft, MCP SEP, drift dashboard (in [STATIC DEMO] mode by default; configure a backend to make it live, see below).
Locally, with a real backend:
bun install
PORT=4300 SEED=1 DB_PATH=/tmp/corpus.sqlite bun apps/corpus/src/server.ts & # observatory + drift dashboard API
PORT=4200 DB_PATH=/tmp/agg.sqlite bun apps/aggregator/src/server.ts & # signed-feed search engine
PORT=4100 bun apps/web/src/server.ts & # marketing site
open http://localhost:4100You'll see:
http://localhost:4100/— homepage with a real cross-source divergence in the herohttp://localhost:4100/dashboard.html— paste any GitHub repo or origin → see all observations + highlighted divergences + a draft signed-feed entry to publishhttp://localhost:4100/spec.html— RFC-style SPEC viewer with sticky TOChttp://localhost:4100/docs.html— CLI / library / adapter quickstarthttp://localhost:4100/search.html— signed-feed aggregator (honest empty state until publishers exist)
The corpus seeds itself from real public sources on first boot: 500 servers from the official MCP registry, 50 agents from the A2A registry, 2,200+ entries from the awesome-mcp-servers GitHub catalog. After it warms up, paste https://github.com/Auctalis/nocturnusai into the dashboard — it'll show you the same server described two different ways by two different sources.
| Status | |
|---|---|
| Spec (v0) — SPEC.md | 1,190 lines · reader contract before producer schema |
| Library — TypeScript reference | src/ · canonicalize · sign/verify · build/parse · Reader |
CLI — agent-feed init / sign / verify / lint / snapshot |
src/cli.ts |
| Publisher adapters | Next.js · FastAPI (cross-language verified) · Cloudflare Worker |
| Aggregator — search engine for signed feeds | apps/aggregator · SQLite + FTS5 + REST + UI |
| Corpus observatory — third-party drift detection | apps/corpus · MCP registry + A2A registry + GitHub READMEs · cross-source divergence as headline |
| Web — homepage + dashboard + spec + docs | apps/web · 5 pages from a Claude design handoff |
| IETF draft | docs/IETF-DRAFT.md · 2,691 lines · draft-abdi-agent-feed-00 |
MCP SEP — agent-feed as the streaming layer beneath Server Cards |
docs/MCP-SEP-agent-feed.md |
| 78 conformance tests across 13 files | bun test |
Every origin gets two artifacts at two URLs:
| URL | Mutability | Answers |
|---|---|---|
/.well-known/agent-card.json |
mutable, replaced | "what is true now" |
/.well-known/agent-feed.xml |
append-only | "how did we get here, and what changed" |
You cannot reconstruct a stream from samples of state. They are different epistemic objects. Mixing them is why migrations break agents silently.
Every fact in this system is one of two kinds:
- Signed — first-party,
did:web-rooted, Ed25519-attested. Testimony. - Observed — third-party, sourced, always tagged
UNSIGNED. Photograph.
The protocol is signed-only. The corpus observatory is observed-only. Storage may join, surfaces never blur. Every observation row carries source + sourceFetchedFrom + observedAt. The dashboard renders unsigned tags loudly, in red, on every observation card. Differently-shaped facts; not a quality gradient.
agent ↔ agent MCP · A2A · ANP how agents talk
agent ↔ api OpenAPI · agents.json how agents act
agent ↔ context llms.txt how agents read
agent ↔ change agent-feed how the world announces it moved ← this project
agent-feed is the change-history layer beneath those protocols. Bring your own MCP / A2A / ANP / OpenAPI; we announce when they mutate. We are not trying to be the agentic web. The web doesn't live in DNS; we don't either.
import { Reader, withFeedRecovery } from "agent-feed";
const reader = new Reader();
reader.on("mismatch", (e) => console.warn("schema mismatch", e));
reader.on("unverified-entry", (e) => console.warn("bad signature", e));
reader.on("feed-migrated", (e) => console.log("publisher moved", e.migratedTo));
await reader.ingest({ origin, xml, didDocument });
// after every live API call, hand the response to the reader so it can detect drift
reader.observeLiveResponse({ origin, endpointId: "orders-api", body });
// the 404-killer: when an agent gets a 404, consult the feed for a replacement
const fetch2 = withFeedRecovery(reader, fetch, { origin });
const res = await fetch2("https://api.example.com/v1/orders");# Generate keypair + did.json + empty feed
bun src/cli.ts init -o https://example.com -d ./public/.well-known
# Append a signed schema-change entry
bun src/cli.ts sign -d ./public/.well-known -t schema-change -p '{
"effective-at": "2026-04-27T13:00:00Z",
"endpoint-id": "orders-api",
"from-version": "1.0",
"migration": { "add": ["currency"] },
"to-version": "1.1"
}'
# Verify and lint a remote feed
bun src/cli.ts verify -o https://example.com
bun src/cli.ts lint -o https://example.combash scripts/full-demo.shPrints a step-by-step walk: agent calls a v1.0 endpoint successfully → origin migrates to v1.1 and signs a schema-change entry → agent re-ingests the feed, reads the migration, and survives the breaking change without redeployment.
agent-feed/
├── SPEC.md # the v0 protocol — 1,190 lines, reader contract first
├── ROADMAP.md # tiered, every item carries its kill-criterion
├── CHANGELOG.md # shipped items moved here per the roadmap discipline
├── docs/
│ ├── IETF-DRAFT.md # draft-abdi-agent-feed-00
│ ├── MCP-SEP-agent-feed.md # snapshot ≠ stream proposal
│ └── design-brief.md # design handoff brief
├── src/ # reference TypeScript library + CLI
├── tests/ # 78 tests + conformance vectors
├── adapters/
│ ├── next/ # @agent-feed/next (44 LOC)
│ ├── fastapi/ # agent-feed-fastapi (Python; cross-language verified)
│ └── cloudflare-worker/ # @agent-feed/cloudflare-worker
├── apps/
│ ├── aggregator/ # signed-feed search engine — SQLite + FTS5
│ ├── corpus/ # observatory — MCP registry + A2A + READMEs
│ └── web/ # homepage + dashboard + spec + docs (5 pages)
├── examples/ # publisher fixture + consumer-survives-schema-change demo
├── scripts/
│ ├── full-demo.sh # one-command end-to-end demo
│ └── build-vectors.ts # regenerate conformance vectors
└── design materials/ # Claude-design handoff: source HTML/CSS for apps/web/
See ROADMAP.md — four tiers, every item has explicit kill-criteria. Shipped items move to CHANGELOG.md per the roadmap discipline. Highlights:
- Tier 1 (close the loop): WebSub push, did:web key rotation, HTTP
Cache-Controlhonoring. - Tier 2 (adoption): hosted conformance checker site, npm/JSR publish.
- Tier 3 (standards posture): IETF dialogue, MCP TC engagement, A2A extension, Rust + Go ports, formal TLA+ model.
- Tier 4 (the bets): the 30-year archive, the regulatory wedge, IDE/browser integration, ML training corpus, time-travel debugging — all gated on real-world signal.
- Not a registry. No central index. Discovery is by origin URL.
- Not a discovery protocol. MCP Server Cards / A2A Agent Cards already cover state-snapshot discovery; agent-feed is the temporal-history layer beneath them.
- Not a status page. Operational telemetry has different time-constants and consumers.
- Not a policy engine. Pricing, rate limits, ToS belong in a separate slow-changing document.
- Not the agentic web. It's a layer, not a runtime. The web doesn't live in DNS; this doesn't either.
The web app is fully static and lives on GitHub Pages. The drift dashboard's live mode requires the corpus app reachable somewhere on the public internet. Two ready-to-go paths:
# Fly.io (recommended — Bun-native, persistent volume for SQLite)
fly launch --no-deploy --copy-config # uses apps/corpus/fly.toml
fly volumes create corpus_data --size 1
fly deploy
# Or Docker anywhere
docker build -f apps/corpus/Dockerfile -t agent-feed-corpus .
docker run -p 4300:4300 -v $PWD/data:/data agent-feed-corpusAfter deploy, point the dashboard at it by setting window.__CORPUS_ORIGIN__ = "https://your-corpus-host" in a <script> tag before the dashboard's existing inline script. The live [STATIC DEMO] banner disappears and the form starts hitting the real API.
Independent. MIT-licensed. The two architecture roundtables that produced the trust-plane separation (pg+carmack+taleb+hickey for the protocol; hickey+carmack+taleb+karpathy for the corpus) are committed to this repo as design provenance. No company. No funding. No signup. agent-feed.dev, not app.agent-feed.dev.
MIT.