What has been verified
These pages describe Plompy Local 0.13.0. The release provides separate core and browser images plus a downloadable starter kit. The older v0.12.0 Discord image does not contain Local mode. A container release does not establish that your model account or third-party API credentials work.
Version boundaries#
| Component | Contract |
|---|---|
| Local runtime | Node 26.13.0 or newer Node 26 |
| MCP SDK | Pinned by packages/local-mcp/package-lock.json |
| Docker | 0.13.0-local and 0.13.0-local-browser, Linux amd64 and arm64 |
| Host clients | Codex and Claude Code export; explicit OpenCode V1 or V2 config |
| Local inference | Ollama configuration guide, not tested in this acceptance run |
| Existing engine OpenCode runner | Independently pinned to 1.19.5 |
| Browser | Pinned optional Playwright dependency, sandbox enabled |
| Separate Discord engine image | 0.13.0, Codex runner inside the engine |
Verification levels#
- Offline: synthetic credentials, local fixtures and controlled responses. This can verify validation and error handling without proving vendor access.
- Container: the actual built image starts, authenticates an MCP client and performs the checked local operations.
- Real browser: Chromium runs inside the browser image and reads a controlled fixture with sandboxing enabled.
- Real OpenCode client: the actual client discovers or calls the gateway. A deterministic model fixture is not a real model-provider acceptance test.
- Live provider: an explicitly authorized call to a real account succeeds. Record provider, tool, date and outcome without secrets.
Acceptance record#
The following checks were completed on 12 November 2026. The test sources and reproducible commands are in the local MCP test guide.
| Check | Result and boundary |
|---|---|
| Core and browser Docker images | Both source targets built successfully and ran the recorded acceptance checks. Release images and architectures are verified separately by the release workflow. |
| Real sandboxed browser | Chromium ran as UID 1000 with no sandbox-disabling flag. It read JavaScript-rendered contacts, retained the final source URL, and blocked POST and metadata requests. |
| Actual Codex 0.159.2 client | An isolated client parsed the generated TOML and completed plompy_status against Docker. A deterministic local Responses mock handled two model requests; a completed MCP tool event was asserted. |
| Actual Claude Code 2.1.197 client | An isolated client completed an authenticated MCP connection health check and reported Connected. No Claude agent/model turn was run. |
| Actual OpenCode 1.19.5 client | An isolated client completed eight real MCP calls against Docker for a synthetic research task. A deterministic local mock model directed the calls. |
| Research artifacts | Three controlled source pages produced two deduplicated company records, draft content, JSON and CSV. Stored files and formula-safe CSV were checked. |
| Container restart | Configuration, gateway token and exact artifact bytes persisted across a real restart. |
| Public OpenAI Docs | Real remote initialization and search_openai_docs with query API Responses and limit 1 succeeded at 15:31 UTC. No model account or paid inference was used. |
| Other paid API providers | Not live-tested with account credentials. Offline validation does not prove provider access. |
| OpenCode V2 | Configuration generation is supported; the V2 client was not exercised in the recorded acceptance run. |
| Ollama | Documented setup only. Ollama and local-model quality were not tested, and no model was downloaded. |
A configuration export or CLI registration alone must not be described as a completed model-powered task. The Codex and OpenCode checks used actual clients with controlled model fixtures; the Claude Code check covered authenticated server connectivity only.
The synthetic research test proves client transport, tool execution, rendering and persistence. It does not prove live search-provider coverage, paid model reasoning quality, mailbox access or email deliverability. Third-party integration failures remain possible because of authentication, plan limits, changed tool names or service availability.
Deliberate first-version limits#
- No email sending, mailbox integration or bulk outreach.
- No nested agent or model inside the gateway.
- No automatic import of host credentials or browser profiles.
- No vendor OAuth implementation in the gateway.
- No Camoufox Docker support in this profile.
- No TinyFish autonomous task tool in the research profile.
- No arbitrary custom MCP servers in the local research profile.
- No generic shell, arbitrary browser script or unrestricted tool-invocation endpoint.
- Documentation content is English; the existing landing retains English and Polish routes.