Shawn Core MCP
# Shawn Core MCP
Shawn Core MCP is a local-first JSON-RPC MCP server for practical specialist
routing, continuity, and evidence-based handoffs.
The public orchestrator is named **Orchestrator**. It retains a concise map of
every specialist for ordinary tasks and escalates only when deeper proof is
needed. That keeps routing helpful without turning every small request into a
multi-agent ceremony.
Every public specialist has the same 8/10 contract baseline: intake, evidence,
handoff, and public boundary. This measures public operational completeness,
not a claim that every domain has identical technical depth.
## Included core tools
- `orchestrator_route`: return the first specialist and validation plan.
- `token_master_audit`: inventory token-relevant local configuration facts.
- `changelog_intake`: start a lightweight changelog subprocess plan with an
Atlas Cartographer handoff, a local Git snapshot, and an `@github` handoff.
- `changelog_finalize`: render a receipt-backed changelog entry at task end.
- `atlas_continuity_record`: write a compact, payload-free Cartographer card
beneath an operator-selected local continuity root.
- `personal_git_snapshot`: read status and latest commit from an allowlisted
local Git repository.
- `personal_git_commit`: make an explicit local commit only within roots
allowlisted by `SHAWN_CORE_GIT_ROOTS`.
The Changelog specialist is configured for `gpt-5.6-luna`, a 64,000-token
compaction target, and a 100,000-token hard ceiling. Those are an execution
contract for the host that runs the subprocess, not an attempt to claim that
an MCP can raise a provider's actual context window.
## Run locally
```powershell
python -m unittest discover -s tests -v
$env:SHAWN_CORE_GIT_ROOTS = "<local-repository-root>"
python -m shawn_core_mcp.server
```
Send newline-delimited JSON-RPC requests over standard input. The server is
local only. It does not open a port, read credentials, change model settings,
or publish to GitHub.
## Public boundary
This repository intentionally excludes personal configuration, model and
dataset paths, cloud credentials, client material, generated outputs, and
private specialist implementations. GitHub publication is performed by the
operator, not by the MCP.
TDQS
Scored across 8 tools
Each tool targets a clearly distinct function: orchestration context vs routing, git snapshot vs commit, changelog intake vs finalize, plus audit and continuity recording. There is no meaningful overlap or ambiguity between the tools.
All names use snake_case with useful domain prefixes, but the second element mixes nouns and verbs: orchestrator_context, changelog_intake, and personal_git_snapshot are noun-style, while orchestrator_route and changelog_finalize are verb-style. The pattern is readable but not consistently verb_noun.
Eight tools is a well-scoped size for the apparent domain of personal orchestration, git snapshots/commits, changelog handling, and continuity tracking. Each tool represents a meaningful capability without excessive fragmentation.
The main workflows are covered: routing and context for orchestration, git snapshot and commit for local changes, changelog intake and finalize for task closure, and a continuity record. Minor gaps exist around updating/reading registry or continuity artifacts, but core workflows have no dead ends.