Operations — run, review, and inspect surfaces.

Five first-party MCP servers behind one composite dispatcher.

PeptiterMCPCatalog.makeComposite(...) returns a CompositeMCPServer routing literature/* → LiteratureMCPServer (4 adapters: fixture, PubMed, Semantic Scholar, Europe PMC), lean/* → LeanVerifierMCPServer (TensorLang manifest hash binding + lake build round-trip), lab_loop/* → LabLoopMCPServer (V3LabLoopOrchestrator behind a typed MCP surface), improvement/* → ImprovementMCPServer (recursive evidence improvement proposals), model/* → PeptiterModelMCPServer (semantic_hash-pinned biology_v* introspection plus world-model source, pathway assertion, and body-twin model-card inspection). Same code path drives the in-process app, the stdio binary, and Swift tests — they cannot drift.

Newline-JSON-RPC stdio server, single binary, ready for Claude Code.

Drop-in for any MCP client that speaks newline-delimited JSON-RPC over stdio. .claude/mcp.example.json + docs/MCP_CONFIG.md show the wiring with three flavors (built-binary path, swift-run dev, concurrent dispatcher). Runs sequentially by default; --concurrent N opts into a TaskGroup-backed pool with an actor-serialized stdout writer so JSON lines never interleave at the byte level.

peptiter-mcp-stdio --probe                  # tools/list as JSON, exit
peptiter-mcp-stdio --tools-list-version     # SHA256 of canonical catalog
peptiter-mcp-stdio --concurrent 4           # batch-friendly dispatch
peptiter-mcp-stdio --cache-dir ~/peptiter   # persist trace + verifier receipts

Lake builds and orchestrator runs persist across restarts.

OnDiskCache<T: Codable & Sendable> generalizes the persistence pattern: actor-backed in-memory dict, optional <dir>/<name>.json file per entry, sanitized filenames, corruption-tolerant load. lean/verify caches receipts keyed by manifest semantic hash so identical-artifact requests skip lake build (or pass noCache: true to force a fresh verify). lab_loop/run caches Trace + fingerprint so cold restarts answer lab_loop/inspect without re-dispatching the orchestrator. improvement/* keeps proposal state inside the MCP server instance; persistence can be added with the same cache pattern when promotion records need to survive process restarts. --prune-cache 7d evicts at startup; peptiter-mcp-cache --evict <ns>/<key> evicts a single entry on demand.

cachedHash and cachedFingerprint let agents skip identical work.

Every expensive tool accepts an optional cache hint from the caller. lean/verify takes cachedHash; lab_loop/run and lab_loop/inspect take cachedFingerprint (FNV-1a over attribution_hash + overrides). Match → compact { unchanged: true, … } delta. Mismatch → full receipt with a delta block flagging the change reason. Stateless on the wire — caller owns the cache.

peptiter-mcp-cache reads what's persisted without spinning up the server.

Read-only inspector. Walks the cache directory, reports per-namespace counts and bytes, lists each entry's key + size + age + fingerprint. --json mode emits a stable schema; --details prints the first 160 chars of each payload (text) or the full decoded JSON. Lets operators audit deployment state offline.

peptiter-mcp-cache --cache-dir ~/peptiter
  peptiter cache @ ~/peptiter
  ## lab_loop (1 entries, 11.3 KB)
    v3-il23-axis · 11538 bytes · 0s old · fingerprint 14ebfd30a3a9e488
  ## lean (3 entries, 4.7 KB)
    biology_v2__e065d18796a17c58… · 1820 bytes · 12m old

Every PR posts an EVAL.md badge; overlay drift fails the build.

.github/workflows/eval-badge.yml runs the transpiler Python tests, regenerates BiologyV2/V3.lean from the (potentially modified) overlay JSON, runs lake build, executes the eval harness, and posts a sticky PR comment with the headline F1. check_overlay_sync.py blocks PRs where BIOLOGY_OVERLAY.md drifts from build_biology_v3. EVAL.md gains a 'Diff vs base ref' section via --diff-against-base origin/main so reviewers see cumulative regression.

EVAL · v2 signed — F1 0.818 · safety precision 1.000 · 25 cells, 4 models