Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@vercy search the catalogue for a model about memory governance"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Two tools for agent memory, in one package with no dependencies.
vercy check - checks a memory store against the Vercy Governance Overlay:
the fields a store needs to answer questions about time, authority and disclosure.
vercy mcp - a read-only MCP server over the published Vercy model catalogue.
Apache-2.0. No telemetry. The checker makes no network calls. The MCP server only sends GET requests to
one origin, https://ver.cy by default (VERCY_BASE_URL to use a mirror), and refuses any URL or
redirect outside it.
uvx vercy check memory.jsonlInput is a JSON array, or one JSON object per line. Your field names can stay your own:
uvx vercy check graph.jsonl --map record_id=uuid,valid_from=valid_at,valid_to=invalid_at,source=episoderecords: 30
level reached: 1 of 3
profile: vercy-governance-overlay 1.0
concept_owner level 2 MUST D-101, D-102, D-103, D-201, D-202, D-301 and 24 more
conflict_policy level 2 MUST no conflict policy anywhere in the storeLevel | Fields | What the fields let a host do, if it uses them |
1 |
| Answers about any date; an update can name what it replaces |
2 |
| Disagreement resolved by rule, change requests routed to an owner |
3 |
| What may cross a boundary, without leaking or over-refusing |
Options: --json for a machine report, --min-level 2 to raise the bar, --policy when the conflict
policy lives outside the records. Exit codes: 0 the store reaches the level, 1 it does not,
2 the input could not be read. That makes it a CI gate.
What a pass means. The overlay fields are present and not empty (valid_to only has to be
present: null means still open). Nothing more. Values are not
validated, and a pass does not mean a host enforces them: a record can carry release_to while another retrieval path ignores it.
Enforcement is a separate, testable claim, defined in ENFORCEMENT-CONTRACT.md.
Related MCP server: TokenLab MCP Server
claude mcp add vercy -- uvx vercy mcpOr in any MCP client configuration:
{ "mcpServers": { "vercy": { "command": "uvx", "args": ["vercy", "mcp"] } } }Tool | What it returns |
| Ranked models for a name or need |
| One model by id, model id, slug or unique alias; with |
| The overlay fields, levels, definitions and measured effects |
| The |
| Citation text, page URL, specification URL, version and digest |
All tools are read-only. Tool failures come back as {"error": code, "message": ...} with codes an
agent can branch on: unknown_model, not_published, ambiguous_id, upstream_unreachable,
foreign_url. Malformed calls get JSON-RPC errors with the request id kept.
Each field is in the profile because a published benchmark measured what changes without it. On the collaborative-memory benchmark, a bitemporal graph without owner and conflict rule answered 77.4 percent. Adding a governance catalogue to the same graph took it to 93.5 percent (p = 0.039). A team wiki and a Vercy dimension carrying the same facts with the fields scored 96.8 and 95.2 percent. Methods, raw runs and harnesses: ver.cy/benchmarks/collaborative-memory-v1.
python -m unittest discover -s testssrc/vercy/data/profile.json is generated from profile.yaml (the file published at
https://ver.cy/overlay/profile.yaml) by python tools/sync_profile.py.
This server cannot be deployed
Guarded MCP server for agent-readable business truth, provenance, readiness, and discovery.
Read-only MCP server exposing a user ORANO library to their own AI agent.
OAuth-protected, read-only-by-default MCP server for provenance-labeled QuillCaddie project memory.
Public, read-only MCP server for FarmNeural company facts, packages, and capabilities.
