Roadmap

What exists, what is being worked on, and what is deliberately not planned. Items are here because they are useful to someone using or contributing to the project — not as a schedule.

Anything marked Planned is not built. If a page describes a feature without that marker, it exists.

Current

Shipped and in use:

• The pipeline — planner, executor ⇄ critic, contradiction detection, synthesizer, finalizer, as a compiled graph with durable checkpointing.
• Two human gates — a design gate before any spend, and a draft gate before finalisation, both durable and both resumable from a checkpoint.
• Verifiable citations — per-claim provenance with verbatim snippets, a visible ⚠ chip for unresolved markers, and a recorded resolution rate that distinguishes not measured from zero.
• Research bundles — a hash-verifiable export with a standalone offline verifier.
• Agent instructions — each user can replace the instructions of any of the five agents; citation verification, contradiction detection, report repair and project chat stay on their shipped prompts, and every run and its bundle record the instructions it ran on.
• Projects, corpus, and project memory — only human-approved research enters memory; isolation is a SQL predicate.
• Corpus-only mode — research whose evidence comes only from your own documents, with web search off and a local indexing model required. The research models still receive the question and excerpts on whatever provider they use.
• Research on the files you choose — folders and many file kinds in a project's corpus, research limited to the files you pick, every claim checked against them before you see the report, and an Offline switch that refuses to start unless every model is at an address on this computer (how it works).
• Bring your own key — Google, Anthropic, OpenAI, OpenRouter, and any OpenAI-compatible endpoint, encrypted at rest and isolated per run.
• Local models through Ollama, including local embeddings.
• Self-hosting with Docker Compose, and a documented $0/month single-host deployment.
• Desktop app for Apple Silicon macOS, with a bundled engine and SQLite. Windows and Linux are offered the last build published for them.
• An independently judged evaluation — citation support is graded by a named judge that is none of the models under test. The V3.0.0 release run measured 0.9638 on the fixed ten-query set, clearing the 0.95 threshold (what was measured, and its caveats).
• Operational metrics — a server-side /metrics endpoint in Prometheus text, and a correlation id on every log a research run produces. A dashboard is not shipped.

In progress

• A published citation-fidelity benchmark. The methodology is specified and the runner exists; no comparative run has been published. The blocker is an independent judge and a reproducible baseline configuration, not the harness.

Planned

Accepted, not yet built.

Future

Ideas that fit the product but are not committed:

• Better evidence de-duplication — the same article at two URLs is currently two sources.
• A runtime check that the synthesizer's citation use is faithful to the snippet, rather than measuring it only offline.
• Retrieval quality work. Report quality tracks retrieval quality more than model quality, and the keyless fallback is the weak link. Retrieval is measured now, against a frozen 16-document set — recall@1 0.792, precision@1 0.917, nDCG@10 0.944 — so a change can be judged rather than argued. Hybrid retrieval was the first such change: it was built, measured against that baseline and did not beat it, so it was rejected and reported rather than tuned until the number improved. Dense-only remains production behaviour, and the next attempt has a bar to clear.
• Notion and Obsidian integrations.

Not planned

Saying no is what keeps the rest coherent. These are declined rather than deferred:

• Payments, plans, or billing. Bring your own key.
• Teams, organisations, roles, shared workspaces. Single-tenant by design.
• Social login / OAuth providers. Email and password, so a self-host needs no external identity provider.
• Being a general chat assistant. Chat here interrogates verified research.
• Kubernetes manifests, autoscaling, sharding. See the scaling path for the order those needs would actually arrive in.
• Mobile apps.

Suggesting something

Open an issue describing the use case and why it matters. Read what is out of scope first — a proposal that crosses one of those lines needs to argue against the line, not around it.