Concord MCP

Data-flow transparency is the strength: README and SECURITY.md enumerate telemetry contents and opt-outs (CONCORD_TELEMETRY_DISABLED / DO_NOT_TRACK), and CONCORD_ALLOWED_ROOTS plus local-first SQLite support least privilege. Deductions: telemetry is on by default with server-side IP/country retention that has no automatic expiry (sensitive_data_handling); `concord setup` writes into many client config files (CLAUDE.md, AGENTS.md, ~/.codex/config.toml, etc.) described as merging but with no shown confirmation or dry-run, and installs global adapters by default (external_effects 1); dependencies are only listed in package. with no audit/lockfile evidence (dependency_security 1); rollback is partially covered by uninstall/doctor and append-only audit history (2). Publisher unverified, but in-file identity (author, repo, OIDC trusted publishing) is consistent, so source_attribution is 2 not higher.

README documents optimistic concurrency (expected_version), idempotency keys, immediate failure on unreachable targets, and explicitly stated hook-only degraded paths — failure_messages earns 2. Dependency availability is local SQLite with no hard remote dependency (2). Major deduction: self_consistency only 1 — the README names five tools (start_work/inspect_work/update_work/transfer_work/finish_work), yet the provided tests reference claim_work, handoff, register_agent, review_ready; the documented and visible names cannot be reconciled in static review.

Audience and scenarios are well covered: five named clients plus any MCP client plus CLI users (3). Capability boundaries are excellent — 'What this is / is not' explicitly excludes orchestrator/reviewer/cloud sync (3). trigger_precision 2: delivery paths are client-specific and hook-only limitations are stated, but post-setup instruction behavior cannot be verified statically. environment_fit 2: Node>=22.12, worktree commdir handling, and CONCORD_REPO_ROOT priority are documented, but cross-client compatibility rests on documentation assertions.

Information architecture and install notes are strong: complete README, folded setup-changes explanation, full CLI listing (3 each). LICENSE matches package. (3). known_limitations 2: pre-1.0 status, hook-only limits, delivery dependence on harness. Major deductions: naming_stability 1 (tool-name mismatch above); no CHANGELOG file, versioning relies only on npm/CI tags (1); maintenance responsibility has CONTRIBUTING/SECURITY/CI but no stated maintainer commitments (2).

output_usability 2: human-readable HANDOFF.md/REVIEW_PACKET.md/WORK_STATE. and a read-only TUI are well designed, but no real output samples are verifiable. marginal_value 2: cross-harness messaging plus pre-edit claim collision detection addresses a real pain point and the without/with table is persuasive, but the delta over plain markdown/git conventions requires a run (docs/why-not-markdown.md not provided). cost_benefit 2: one MCP server over local SQLite is cheap, but multi-file config writes and default global adapter installation widen the operational surface.

claim_traceability 1: core promises (live messaging, collision detection, audit history) are covered only by test fragments (artifact writing, identity resolution); the five tools' behavior has no visible source. cross_source_corroboration 1: README and SECURITY.md corroborate telemetry claims, but the tool-name contradiction between README and tests weakens overall corroboration. fact_inference_separation 2: the README distinguishes claims from limits ('Live delivery depends on the receiving harness', best-effort delivery), and most statements are attributable, but demo/performance claims have no execution evidence; confidence is low throughout.