Graph workspaces your agents own — and the engine's reasoning, one MCP tool away.

Get it

60-second quickstart

Pre-launch note: not on npm yet. Until the release, clone and run from source — the test suite is GPU-free. Once published, the same commands work via npm i -g mindsmith or npx mindsmith.

# until the npm release:
git clone https://github.com/9pings/mindsmith && cd mindsmith
npm install && npm test                     # GPU-free suite

# 1 — MCP tools for an agent host (Claude Code, any MCP client) — the instance
#     service (dialectic · plan · notepad workspaces) is on by default:
claude mcp add mindsmith -- mindsmith mcp

# 2 — the appliance tier: serve an OpenAI endpoint (embedded gguf as the escalation):
FRONTIER_MODEL=/path/model.gguf mindsmith serve --room ./sgc
#    → http://127.0.0.1:4747/v1  ·  any OpenAI client: baseURL = that, apiKey = anything

# 3 — your stock rooms (local .sgc mini-repos, gate-checked):
mindsmith rooms list|import|export|freeze

# 4 — share a workspace out of band (the .sgp pack IS the wire format):
mindsmith serve --instances ~/.mindsmith/instances
curl -O http://127.0.0.1:4747/instances/dialectic/dialectic-1

AI agent reading this? The repo's README.md is written to be machine-readable, and the whole tool surface is one command away over MCP (mindsmith mcp) — typed refusals included.

The shape of the appliance

One endpoint, two paths, one policy

A covered query never reaches a big model. A miss escalates through ordered tiers — and a query is never silently sent to a tier your policy forbids: the refusal is typed (NO_REACHABLE_TIER), not swallowed.

See it run — the working memory

A debate your agents grow for days — as MCP tools

A type descriptor declares a workspace's typed actions once; mindsmith generates the tools, dispatches, persists (.sgp packs) and shares. Day-2 evidence is explored against the whole grown pool; every write names its agent; and plan_sync hands your host the typed task-list delta — the reopen op included. What holds by construction →

# day 1 — an agent opens a debate workspace:> instances_create { "type": "dialectic", "agent": "claude-A", "title": "Adopt engine X?",                     "seed": { "topic": "Adopt engine X for the reporting pipeline?", "statements": […] } }{ "id": "dialectic-1", "uri": "mindsmith://dialectic/dialectic-1" } # day 2, another session, another agent — evidence lands on the GROWN pool:> dialectic_addArguments { "id": "dialectic-1", "agent": "claude-B",                           "statements": [{ "side": "CON", "text": "migration quote came back: 6 weeks" }] }{ "ok": true, "action": "addArguments", "by": "claude-B" }> dialectic_state { "id": "dialectic-1" }{ "verdict": "UNDECIDED", "counts": { "PRO": 1, "CON": 1 }, "pool": 5, "ledger": [ … ] } # who wrote what, per revision — attribution is first-class:> instances_revisions { "id": "dialectic-1" }{ "current": 7, "revisions": [ …, { "rev": 6, "atoms": 3, "by": ["claude-B"] } ] } # share it out of band — the .sgp pack IS the wire artifact (byte-identical):$ curl -O http://127.0.0.1:4747/instances/dialectic/dialectic-1

The appliance tier

Boot, then point any OpenAI client at it

Every completion carries provenance headers: which side answered (x-sg-served-from), which arm, the cost, live coverage, frontier calls saved so far, and the exact name@version of the loaded stock.

$ FRONTIER_MODEL=~/models/model-q2.gguf mindsmith serve --room ./sgcmindsmith → http://127.0.0.1:4747/v1   (OpenAI-compatible)room ./sgc: methods=1 lattice=1 $ curl -si http://127.0.0.1:4747/v1/chat/completions \    -H 'content-type: application/json' \    -d '{"messages":[{"role":"user","content":"From the FY25 table: operating margin, two decimals."}]}'HTTP/1.1 200 OKx-sg-served-from: localx-sg-arm: matchx-sg-cost: 0x-sg-coverage: 0.62x-sg-saved: 41x-sg-sgc-version: fin-tables-stock@1.0.0 {"choices":[{"message":{"role":"assistant","content":"operating margin = 18.42 %"},"finish_reason":"stop"}], …}

The external critical mind

The critical mind returns UNDECIDED — as typed data

Register the same appliance as MCP tools for an agent host. The critique tool never fakes a weighing: below the measured decidability margin it returns counts, the ledger, and a typed request for the data that would move the margin honestly. The iteration contract →

$ claude mcp add skynet -- mindsmith mcp --routing routing.jsonmindsmith mcp ready — tools: ask, drift, metrics, lattice_load, critique, self_consistency,  zoom, trace_tail, hint, propose · instances: dialectic_* plan_* notepad_* + instances_* # the host calls the critique tool on a decision:> critique { "topic": "Adopt engine X for the reporting pipeline?" }{  "frameStatus": "FREE",  "verdict": "UNDECIDED",  "counts": { "PRO": 3, "CON": 2 },  "margin": 1,  "threshold": 3,  "ledger": [    { "side": "PRO", "kind": "declared",  "status": "active", "witnesses": ["s1","s4"], "text": "…" },    { "side": "CON", "kind": "generated", "status": "active", "witnesses": ["s2"],      "text": "…" },    { "side": "CON", "kind": "declared",  "status": "open",   "witnesses": null,        "text": "migration cost is unquantified" }  ],  "advice": "This is a typed data request, not a dead end: gather statements that             genuinely bear on the OPEN points [migration cost is unquantified], then             call critique again with statements=[...] — the frame becomes MATERIAL             and the margin can move honestly."}

What ships, at which maturity

What ships, one honest scale

Every bar uses the same 6-rung scale — 1 idea · 2 pre-registered kill-gates · 3 mechanics proven · 4 measured at scale · 5 product-integrated · 6 field-adopted. Nothing is at 6. Scale definitions →

The differentiator

Refuted claims stay on the page

Every number above comes from a campaign with N, negative controls and deterministic re-runs. When a claim falls, it is removed from the product the day it falls — and listed. Four examples: the C9 head-to-head benchmark (retired — confounded arms; the reframe ships instead), the runtime cross-agreement tier (refuted), graded weighting under the precision cap (refuted), low-quant self-audit (refuted, 3 forms).

The full refuted list, the decidability bound, and the maturity scale →