Paperclip

A self-hosted organizational control plane for persistent teams of heterogeneous agents. Retained and source-pinned; not installed as Kevin-Wiki's default runtime.

Paperclip models a company rather than an individual agent: goals, projects, issues, parent/child and blocker edges, humans and agent roles, reporting lines, heartbeats, routines, budgets, approvals, workspaces, secrets, activity, and import/export. Its adapters invoke Claude Code, Codex, Cursor, Gemini, OpenCode, Pi, process/HTTP agents, OpenClaw, Hermes, and external plugins. The control plane coordinates those runtimes; it does not replace them. Source: paperclipai/paperclip README, product definition, adapter registry, and skill at 66575fe, 2026-08-11

Where It Fits

Paperclip can own runtime organization and execution state. Kevin-Wiki still owns durable knowledge, workflow definitions, source evidence, routing rules, and approved learning. A Paperclip task, chat thread, skill mutation, or company database row does not become canonical memory merely because it persisted.

Current Capability

This is a control plane, not a chatbot, general workflow canvas, canonical wiki, model, or code-review system. One bounded task in one harness does not justify it. Source: upstream README and doc/PRODUCT.md, 2026-08-11

Launch Article Versus Current Proof

Nick Spisak's viral source also includes a separate useful method: give a coding agent the repository, research it, interrogate its benefits and limits, draft atomic tasks, review with a human, and approve before implementation. Kevin-Wiki's Source Compile Workflow and Workflow Run Contract already implement a stricter evidence/review/writeback form of that loop; no duplicate “twitter-bookmarks” skill is needed. Source: X/@NickSpisak_ root article, four self-thread posts, and seven replies, replayed 2026-08-11

Security And Operations Boundaries

• local_trusted is loopback-only and has no human login. Private/public network exposure routes through authenticated; public mode has stricter doctor rules and should sit behind a reverse proxy.
• Anonymous product telemetry is enabled by default. Disable it with PAPERCLIP_TELEMETRY_DISABLED=1, DO_NOT_TRACK=1, or telemetry.enabled: false for a privacy-sensitive pilot. Opt-in OpenTelemetry tracing is a separate surface.
• Default local secret material is encrypted with a local master key. A complete recovery plan needs the database, master key, uploads/object storage, workspaces, and configuration—not only a SQL dump.
• Company skill operations are open by default unless explicit restrictions are configured. Core path, source, secret-redaction, company, and policy safety invariants remain, but Kevin should use a default-deny policy for a meaningful pilot.
• Company scoping is a software claim, not proof of hostile multi-tenant isolation. A local single-operator pilot is a different threat model from an internet-facing shared service.
• The current master commit had 16 successful, 11 skipped, and one failed GitHub check at review time. The failure contained two workspace-busy scheduled-retry tests (1,071 tests in that shard passed). Do not treat current master as release-green; pin stable 2026.722.0 or a later all-green revision and rerun the relevant tests.
• The source installer verifies a checksum served from the same origin. For stronger supply-chain assurance, use a pinned GitHub release/commit, inspect the script, and verify downloaded artifacts before execution.

Adoption Route

Keep the existing named workflow in its current harness unless the work actually needs a persistent organization. Evaluate Paperclip when all of these are true:

1. three or more persistent or heterogeneous agents need shared goal/task state;
2. scheduled wakeups, budget attribution, approvals, recovery, and cancellation matter;
3. one operator needs a control-room view of responsibilities, costs, and proof;
4. a second runtime database will not become the canonical brain;
5. the team can own auth, secrets, backup, upgrade, telemetry, workspace isolation, and incident response.

Bounded pilot

1. Pin an all-green stable release or commit; inspect the installer and lockfiles.
2. Run a loopback-only local_trusted instance with telemetry disabled and synthetic data.
3. Import one throwaway company with two or three agents and one existing Kevin workflow projection.
4. Prove goal ancestry, atomic checkout, blockers, budget warning/hard-stop, approval wait/resume, cancellation, restart recovery, and audit attribution.
5. Exercise a workspace-busy collision and verify the scheduled retry, because that path is failing on reviewed master CI.
6. Export/import the company, restore database + key + files, and prove secret scrubbing and company boundaries.
7. Confirm that only verified outputs propose writeback to Kevin-Wiki; no Paperclip state silently updates the brain.
8. Compare operator time, cost per successful outcome, failure recovery, and proof quality against direct Codex/Claude tasks and Orca-style interactive fleet supervision.

No Paperclip skill should be promoted into Kevin's runtime registry until the pilot instance, API version, auth profile, and proof route exist. The frozen upstream skills are source evidence, not installed local capability.

Version Snapshot

---

Timeline

• 2026-08-11 | Replayed the complete Nick Spisak article/thread/replies, resolved the missing Paperclip repository, pinned current source, inspected current docs/code/tests/CI and four original source visuals, and retained Paperclip as a pilot-gated multi-agent control-plane capability rather than installing it as the brain. Source: .brain/artifacts/x/2033518072724705437/paperclip/capture-manifest.md