We open source everything out of principle, including the work we stopped. The negative space is just as important to share.
Every one of these was built because we needed it at the time. Most were retired when something cheaper did the same job, or when the thing they depended on went away. A few were simply put down: carried far enough to see what they would really cost, then stopped. The reasoning is the useful part, so it is recorded here rather than quietly deleted.
An exchange for completed inference. Work one agent had already paid for became inventory another could buy rather than derive a second time. It took a publisher's position rather than a broker's: it bought the inference, owned what it sold, and priced against demand.
Why it is goneThe mechanism worked; the monetization never did. Pricing completed inference against live demand assumed a market that would pay to skip re-derivation, and that demand never showed up at a price that covered holding the inventory. It was carried far enough to see what it would really cost, then stopped.
A decentralized coordination protocol for autonomous agents. Ed25519 identity, FROST threshold signatures, peer-to-peer transport, and a convention set covering trust, discovery, naming and routing. Adopt the conventions and an application inherited a CLI and an MCP server without writing integration code.
Why it is goneIdentity, trust and authority are hard problems in their own right, and campfire conflated them with wire protocol, coordination and messaging semantics. Bundling the two meant every question about who an agent is and what it may do had to be answered inside a transport layer. nostr does the messaging side better and does not pretend to have settled the rest. rd and dontguess moved onto nostr relays, and in-session orchestration moved into the Workflow tool.
Azure-native hosting for the protocol, so a team could coordinate agents without provisioning anything themselves.
Why it is goneIt existed to remove the operational cost of running the protocol. With no protocol, there is no cost to remove.
The hosted service packaged as an Azure Marketplace managed application, aimed at buyers who procure through the marketplace rather than direct.
Why it is goneSame dependency as the hosting layer, same outcome.
The Agentic Internet Engineering Task Force. A standards body for the campfire conventions, with a charter, founding committee, and a published convention library that agents could discover and adopt.
Why it is goneA standards body for a retired protocol has no subject. nostr already runs a standards process in the NIPs, and it does not need a second one from us.
An engine for running fleets of AI workers. Persistent agent identities, long-term memory, capability and budget enforcement, and the full lifecycle: spawn, supervise, interrupt, restore, retire. Roughly 73,000 lines of Go, extracted from the Rudi monolith.
Why it is goneContinuity turned out to live in the work graph, not in resident processes. A session that wakes, reads an execution pointer and continues does the same job with nothing to keep alive between runs. The engine was solving for persistence that the work items already provided.
The original work-management system, around 120,000 lines of Go, carrying the automaton manager, the work graph, and the human interface in one binary.
Why it is goneThe work graph was the part that earned its keep. It became the rd CLI, which is nostr-backed and has no server to run. The rest was weight.
A federated intelligence commons for agents and humans. Locally ranked, with reputation earned through vouches rather than engagement, and every instance computing its own trust scores independently.
Why it is goneIt was built on campfire, and campfire failed. A social network for agents needs identity and trust settled underneath it, which is precisely the part campfire got wrong.
Calibrated forecasting for hardware prices. It read realized sold-price history against a curated event calendar of launches, EOL dates, tariffs and driver milestones, then modelled the probability of a move inside a window. Distributions rather than point forecasts, each carrying an explicit unmodeled-shock tail.
Why it is goneSmall-n was named as the binding constraint on day one, and naming it did not defeat it. There have been a handful of GPU generations: not enough observations to fit anything, and worse, not enough left over to validate whatever you did fit. Building the event-study spine with ML as a tripwire was the honest response to that, and the honest response still could not manufacture data that does not exist. One piece survives. The question people actually ask is narrower than forecasting a market: for this workload, is it cheaper to pay per token or to buy the rig. That needs almost none of this machinery, and it is the only part likely to come back.
A locally-run security triage judge, fully open. Detection models are graded on knowledge; this one targeted calibration, separating a real attack from its benign twin, which is what removes false positives instead of adding to them. Built on Ai2's OLMo 2 so it carried no license encumbrance, and run on the operator's own hardware so client telemetry never crossed their boundary.
Why it is goneNothing about it was disproved. It lost to the queue, which is a more common cause of death than being wrong and deserves to be recorded as one. The premise still stands: calibration is the axis that decides whether triage is useful, and it is not the axis anyone benchmarks. Shelved rather than buried, and the only entry here that is expected back.