This is the honest tour, no roadmap and no marketing. Just what runs today, how the pieces fit, and why it's built this way.
Accounts are objects on a plane. They have an x-y position, cluster by signal similarity, and sit on a time axis. You zoom, pan, and annotate. The workspace renders in ASCII, so any model you connect reads it exactly as you do. There's no DOM parsing, no screenshot OCR, and no summarization step in between.
northeast ICP fit × ACV · t: now ·─── low ──── mid ──── high ───· ACV hi | ░ ░ ░ ░ ░ ░ · ● Vertex ACV mid | ░ ░ ░ ░ ● Apex ● Cedar ACV lo | ░ ░ ░ ░ ● Calloway ● Meridian · ░ fog: 0 contact ● touched · t-axis: Apex [t-30: mid/lo] ──── Apex [now: mid/mid +14] 90d rolling · annotate · pan · zoom
Scoring reads public hiring, funding, product, and exec movement signals. Each account carries a current score and a 30-day delta. The delta matters more than the snapshot. A 61 trending +18 is the interesting object, not a static 90.
score Δ30d account ──────── ───── ────────────────── 91 ███ +22 Vertex Systems 82 ██▓ +14 Apex Logistics 61 ██ -03 Meridian Health 34 █ ±00 Calloway Partners
Enrichment pulls what is publicly available and attaches it to the account object: org structure, headcount by function, job postings, product shipping cadence, funding, press. It updates on a 6-hour cycle. Every field has a source stamp and a last-seen timestamp. Nothing is inferred that cannot be cited.
Vertex Systems headcount 142 src: linkedin 3h eng % 38% src: linkedin 3h RevOps hires 30d +2 src: careers 1h last raise Series B $28M src: press 14d product ship weekly src: changelog 6h
Build a sequence from a scored account list. Each touch drafts against the specific signals attached to the account. Replies trace back to which touch, which angle, which account. Attribution is per-object, not aggregated into a cohort.
Apex Logistics: sequence/ae-motion t1 email signal: +2 RevOps hires draft ▸ t2 email angle: tooling gap in QBR draft ▸ t3 li contact: VP Revenue draft ▸ ───────────────────────────────────────────── reply → t2 · 4d · replied to RevOps line
Before a meeting, call prep reads the public record and your own notes on the account. It returns the signals, recent motion, and a suggested angle. Where the signal is thin, it tells you. If there's nothing in the public record, the output stays empty instead of inventing something.
Vertex Systems: VP Revenue, 14:00 Thu last contact 11d ago (email, replied) last movement +2 RevOps hires, 14d active signal expansion motion confirmed angle budget cycle Q2, tooling gap open questions current stack · buying process
The Model Context Protocol server exposes twelve tools against the workspace: query accounts by signal, diff a territory over time, draft call prep from public data, compose a sequence against a scored list, annotate an account, and so on. Any MCP-compatible model can connect. It isn't reading a summary of your CRM; it's operating the workspace directly.
tools/list →
query_accounts territory_diff
generate_call_prep compose_sequence
annotate_account score_snapshot
get_signal_sources list_territories
attribute_reply archive_account
draft_followup set_focus
The topology view renders a territory as a field. Accounts cluster by signal similarity, lines indicate relationship, colour indicates motion. The diff view overlays the state from N days ago against today. You see what moved, not just what exists.
topology: northeast diff: 30d
──────────────────────────────────────
Apex●─────────●Vertex ↑+22
╲ ╲ ╱
● ●Meridian ↓-03
Cedar ╲
●Calloway ±00
──────────────────────────────────────
moved: 3 of 5 stable: 1 cooled: 1
Most tools bolt AI on afterward. A chat panel floats over a CRM that was designed before language models existed, and the model only ever reads a DOM tree or a summarised export. It never sees your workspace, only a description of it.
silvercontext flips that. The workspace is text, so the same ASCII structure you navigate is exactly what the model gets when you ask it something. There's no translation step and nothing lost to summarisation. The architecture is the AI interface rather than a layer on top of it.
The MCP server exposes twelve tools against that structure: query accounts by signal, diff a territory over time, draft call prep from public data, compose a sequence against a scored list, annotate an account. Again, it isn't reading your CRM, it's operating your workspace.
Operators working the same territory see each other's motion live. Handoff without losing context.
The building blocks extracted as a library. Spatial workspace + AI legibility as primitives for any serious sales tool.