dextermRoadmap
The order of battle and the rules that bind it. No dates.
The north star: a Solana-only terminal that is fast, extensible and AI-native, and that actually works. One place for screening, wallet and holder intelligence, AI analysis, voice control and non-custodial trading. The spec is the technical source for what is built; this page is the order and the rules. Nothing here has a date, and nothing on the features page is labelled live unless it is on the app's main branch today.
Order#
- Now, shipped: the screener on real data, the cyberspace look, the data inventory, the Claude analyst, the LARP detector, AI scan mode, paper trading, the popup chart host, coin images, a local history recorder, and a keyless trending board with news and KOL lists.
- Next: wallet and holder intelligence (spiked and measured first), the token and feeds popups, bubble maps as popups.
- Then: integrations and social posts (official APIs, user-authorised access only), the AI suite (analysis packs, narrative board, backtests), Jarvis voice, customization (workspaces, widget API, rules engine, bring-your-own-agent).
- Last and most carefully: non-custodial trading, then order plans and standing mandates, then copy trading (alerts first, then a one-key mirror ticket, then capped automation).
Pillars#
| pillar | what | state |
|---|---|---|
| live data firehose | every useful live source, measured not assumed: chain events, launchpads, pools and candles, Jupiter, DexScreener, RugCheck, social, news, SOL macro | inventory done; keyless path shipped; keyed depth pending |
| wallet and holder intelligence | bubble maps (clusters by funding source, same-slot buys, transfer graph), concentration, fresh-wallet / bundle / sniper / dev share, cabal detection, wallet tracker, smart-money ranking, rug fingerprints | spiked (funder = first SOL inflow signer); labels stay hidden until their precision is measured |
| AI analysis | the analyst plus packs: token and wallet dossiers, narrative detection, a rug-probability model backtested on our own recorded history, batch scoring, watch-rule agents, an eval harness | analyst, skills, knowledge base, case files and scan mode shipped; the local history recorder shipped (the data for backtests); packs and backtests pending |
| customization and flows | themes, workspaces, widget API, keymaps, a rules engine (when X then alert / ticket / ask Claude), scriptable config, shareable packs; your own agents over MCP with per-agent capability tiers | themes and keymaps shipped; bring-your-own-agent shipped (task-1345: MCP agent packs + a loopback agent bus, server-side observe/annotate/propose tiers, labelled proposals, a roster popup and a 41-attempt malicious-agent suite — no agent can arm, sign or raise a cap; per-agent mandates still design-only); workspaces, user themes + keymap overrides, the rules engine and shareable packs shipped 2026-10-12 (task-1358: a morning HUD workspace, a contrast + hue guard on themes, reserved keys refused, rules that alert or propose and never arm, packs with a sha256 lock + ssh signature that never run code); the popup widget API shipped as beta (one example widget, a window only; the in-process plugin sandbox and scriptable config pending) |
| Jarvis voice | push-to-talk, an instant local command path, Claude for analysis, spoken replies, popups; clarifies before it proposes and never guesses | shipped on macOS 2026-10-10 (say + on-device Apple Speech or whisper; proposals only; latency measured); a person's first-launch mic pass still pending; wake word, hands-free confirm tiers and non-macOS speech pending |
| trading machine | non-custodial, fast, caps enforced in the signer | PAPER shipped; the LIVE path (wallet link, separate signer process with caps, decode + binding + simulation with token balances, keypress gate, panic) merged 2026-10-10 as in progress, devnet first; mainnet behind a flag and not yet exercised; order plans and standing mandates merged 2026-10-11 as in progress (task-1355: conditional entries, ladders, take-profit, moon bag, stop, expiry; keypress-armed mandates with hard ceilings; a hash-chained journal; Jarvis and agents propose only; paper by default, the live-wallet fire path per fill not yet wired); the fast capped wallet pending; copy trading merged 2026-10-12 as in progress (task-1360: wallet tracker + alerts, one-key mirror proposal, capped auto-mirror behind the less-safe popup; personal, no custody, no fees; paper by default, mirror-sell not built) |
| integrations and social signal | Phantom as a signing bridge, Login with X, per-token X and news feeds, KOL lists, spam filtering; official APIs only, no scraping | keyless layer shipped: trending board from public lists, publishers' RSS, local KOL lists, spam and injection filtering; the Phantom link bridge merged (devnet probe pending); Login with X (posts) pending |
| cyberspace look | maximal and still practical; trade-moment effects and sound only after a confirmed on-chain result | the look shipped; trade effects wait for trading |
| launch and distribution | this site, a public repo with signed reproducible releases (provenance, checksums, SBOM), third-party scan results on the trust page, a marketing kit | this site; the rest pending Joey's go |
Hard rules#
- It must work. Every source goes through the health registry and degrades visibly; never fake data. Every feature ships with replay-based tests and a live smoke script. A compromised-model test suite stays green. End-to-end latency is measured.
- Money safety. Keys stay local. Models and voice create proposals only; arming needs a physical keypress or a wallet approval. Caps, allow-lists and the panic key live in the signer-side policy engine, never only in the UI. The keypress gate is on by default; opting out needs an explicit "this is less safe" confirmation, is local, logged and reversible, and the signer caps still apply.
- Prompt firewall on all untrusted text: token names, bios, tweets, news, fetched pages.
- Legal flags before anything public: fees or token tie-ins, AI "recommendations", copy-trading for other people, wallet-labelling accuracy and privacy. There is no dexterm token.
What "coming" means here#
A plan. Some of it is designed on paper, some is spiked with measurements, some is built on a branch and reviewed; none of it is a promise or a date, and the app never pretends it is live: a coming feature shows a dash, needs key, or a demo label until the real thing lands.