Document library
ENID
GitHub

HeLa MCP Ecosystem Release Notes

Release history

RELEASE_NOTES.md

Original source language. Download source · All documents

v1.6.2 — Plastid and Receptor dependency refresh

Plastid moves to aa43877 and Receptor to 1b21cb7, adopting the lockfile updates that had accumulated on their default branches. Both carried open advisories against @modelcontextprotocol/sdk; the refreshed locks resolve them, and npm audit is clean for Plastid's runtime dependencies while Receptor reports none.

Plastid's manifest also declared "@modelcontextprotocol/sdk": "latest". That is why its lockfile drifted without a reviewable diff; it is now ^1.32.2. Builds and the existing suites pass — 12 Plastid regressions and 123 Receptor tests — and no cell version changed.

Exposure was limited: the SDK advisories concern HTTP transports, body-size limits, and OAuth flows, while Plastid and Receptor run over stdio. Updating still removes the drift and the noise. Hardware, the Blender bridge, GUI clients, and providers remain separate checks.

v1.6.1 — switch deployment modes in place

./setup.sh mode --prefix PATH --mode MODE converts an existing managed installation to another execution mechanism without a second prefix or a data migration. All five modes are valid targets. Profile, snapshot, workspace, transport, port, and the per-cell data/ directories carry over; host-mode switches reuse packages already built at the pinned revision with a clean worktree, and Docker modes build or drop their image and compose.json accordingly.

The command runs the same verification as a fresh installation and generates a new client fragment instead of editing live client settings. Failed verification restores the previous manifest and writes a pending marker so doctor reports the degraded state. Pending or unmanaged prefixes are refused, HTTP targets require a runtime mode, and re-selecting the current mode is an error.

Enzyme stays at 4.0.0 and the runtime at 1.2.0; every cell pin is unchanged from 1.6.0. Tests cover host-family round-trips, transport guards, pending markers, and Docker switches.

v1.6.0 — consistent Enzyme search input

Pin Enzyme source eb98982. google_search now takes query, matching every other single text search in the cell. The former q argument is removed, so a client that still sends it gets an MCP validation error instead of a silent translation; requests that send both fields are rejected as well. Batch tools keep queries, and provider HTTP parameters are unaffected internally.

Both Enzyme entrypoints share one strict schema: the query is trimmed and must be non-empty, num is an integer from 1 to 10, and unknown fields are refused before any provider call. Enzyme's contract tests cover discovery, dispatch, rejection of obsolete input, and provider failure for the shipped and modular servers; 32 contract tests and 111 total Enzyme tests pass with a clean TypeScript build and no lint errors. Enzyme moves to 4.0.0 because this is a breaking argument change.

The catalog, generated documentation, and profile counts are rebuilt from the new schema. Counts stay at 24 Enzyme tools, so profile totals do not change. Other cells, the runtime, and every snapshot from v1.0.0 to v1.5.3 are untouched.

v1.5.3 — canonical relation ordering

Pin Genome source 9ba1d0e. Relation links now render in one canonical order regardless of database insertion order, preventing relation-order-only conflicts after a lossless repair. All 19 suites / 264 tests pass; the TypeScript build and zero-warning source lint pass. The published 1.5.2 snapshot remains unchanged, as do all other cell/runtime pins and package versions.

v1.5.2 — replicated vault convergence

Pin Genome source c32ae94: replicated notes keep the earliest creation timestamp and latest update timestamp, while incoming observation supersets can retain their order without losing database-only facts. The explicit programmatic repair API archives both originals, respects snapshot/lock checks, and permits rejecting a plan before publication. Normal CLI sync does not choose conflict winners automatically.

Genome's 19 suites / 263 tests and TypeScript build pass, including regressions reproducing the previous timestamp/order loops and validating SQL bindings, JSON rows and tool arguments. Source lint reports zero errors and warnings, covers top-level files and fails on warnings. Package versions remain 3.0.0 for cells and 1.2.0 for runtime; this source snapshot records the exact fixed Genome commit. Other pins and historical snapshots are unchanged. Real vault repair and MCP reconnection require separate, backed-up operational verification; no private vault data is published here.

v1.5.1 — current documentation and bilingual website

Documentation-only refresh: all ten cells remain 3.0.0, and runtime remains 1.2.0. The new source snapshot pins the documentation commits without rewriting v1.0.0–v1.5.0. Plastid and Receptor documentation was merged into their default branches without discarding unrelated dependency updates; the snapshot deliberately pins the documentation-only parents, not those dependency changes.

The EN/ID website now loads generated versions, canonical tool names, and profile metadata. Configuration previews use real repository directories and Codex mcp_servers sections. Source guides cover all eight profiles, cold discovery, optional integrations, safe backup/restore, and the limits of current verification. Generated private configurations may contain supplied credentials; committed examples are regenerated with a sanitized environment.

Documentation checks cover 74 regression cases plus local links/anchors and structured metadata; the CLI matrix covers 80 combinations. Browser verification covers 64 visible selections per language and all 16 pages at desktop/mobile widths. These are documentation/UI checks, not new authenticated ACP, provider, Android, or Blender acceptance. Publication does not restart the runtime or change deferred access controls, credentials, databases, or vault conflicts.

v1.5.0 — on-demand runtime and canonical discovery

Runtime 1.2.0 starts cells only when invoked, shares concurrent startup attempts, and discovers schemas without spawning cells. Revision/build/config-aware caches reject stale entries; legacy/corrupt caches fall back to a validated generated catalog. Unknown calls no longer start unrelated cells. Standalone servers, all eight profiles, and the dev-workspace default are preserved.

The canonical inventory drives config generation, doctor/version/profile checks, schema catalogs, and generated documentation counts. Doctor can verify exact runtime discovery with no cell startup via --runtime. A new immutable v1.5.0 snapshot pins runtime 9c92dbf; older versioned snapshots are unchanged.

Verification: 32 runtime regressions on both Linux architectures, 16 all-profile namespace-mode checks, 80 client configurations, four probe regressions, and four inventory regressions. Cold/warm measurements are recorded, not treated as universal benchmarks. See implementation and verification scope. ACP authentication, external client reloads, gateway retirement, vault conflict review, hardware/provider integrations, and deferred security remain separate work.

Servy rollout verified: public/local discovery exposes 236 tools without starting any cells; the first safe Nucleus call starts only Nucleus. Deployed doctor, catalog, and revision checks pass. Fresh backups and rollback material are retained, and all 633 existing conflict archive files are unchanged. See the rollout evidence.

v1.4.0 — memory safety and runtime verification

The v1.4.0 source snapshot pins the published cell revisions and runtime 1.1.0. Older versioned snapshots are unchanged. latest-stable.json tracks this source snapshot, not a guarantee that every optional integration or GUI session has been exercised.

  • Genome: 19 suites / 242 tests with clean lockfile dependencies, including WAL-safe backups, isolated restore, vault conflicts, repeated imports, and exclusive mission claims.
  • Runtime: 20 isolated fault tests covering hung children, SIGTERM/SIGKILL, actual process exit, one restart, profile/cache boundaries, and shutdown cleanup.
  • Nix: real Node 22 package builds on x86_64-linux and native aarch64-linux, using pinned inputs and fixed-output npm dependencies.
  • Dependencies: clean-install Zod 4 build fixes in Genome and Enzyme; Enzyme's 79 tests and Phenotype's 74 tests pass.
  • Ecosystem: ten isolated standalone MCP handshakes and 80 configuration-generation combinations pass. Codex TOML now uses mcp_servers and is accepted by the native parser. These are not authenticated GUI tests.
  • Discovery: 290 standalone tools with LLM and model-backed agent features disabled; the same isolated dev-workspace runtime exposes 230 tools including hela.pipe. Feature-enabled deployments can expose additional Mitosis tools and the optional agent alias; query tools/list rather than assuming a static total.

Known limits: Zed/Codex ACP authenticated-session verification remains pending; duplicate ownership warnings do not establish a connection failure. Existing live vault conflicts remain archived for deliberate review. Android/Blender discovery is not hardware integration. Endpoint authentication and credential rotation remain operator work, not changes included in this release.

Servy deployment verified (2026-11-06): the rooted ARM Nix package serves the eight-cell dev-workspace profile under systemd. Local/public discovery returns 236 live tools, versus 230 in feature-disabled probes; all three Genome mission tools are present. Deployed doctor and strict version/pin checks pass. Existing conflict archives and unrelated client processes were preserved. See deployment evidence and remaining client/gateway work.

All-profile follow-up (2026-11-06): all eight profiles pass runtime startup/discovery, exact tool/schema parity, and safe read-only routing in both namespacing modes (16/16 checks). All ten standalone cells plus runtime pass strict version/pin verification; the 80 client-configuration combinations and four probe regressions pass. Machine selections and release snapshots are unchanged. See profile verification.

Historical v1.0.0 notes

We are proud to announce the official v1.0.0 release of the HeLa MCP Ecosystem — a modular, immortal cellular AI tooling stack built on the Model Context Protocol (MCP), engineered by Azzar.

Architectural Vision & Metaphor

The HeLa MCP Ecosystem is named in respectful recognition of Henrietta Lacks (1920–1951) and the enduring scientific legacy of HeLa cells. Using a cellular biology metaphor, the ecosystem organizes 10 specialized MCP servers into a harmonious organism:

  • Cellular Specialization: Distinct organelles (Nucleus, Membrane, Enzyme, Ribosome, Plastid) perform targeted domains.
  • Deterministic Immortality: Environments can be snapshot-pinned (config/snapshots/v1.0.0.json) and restored cleanly on any machine.
  • Repository Independence: Underlying source repositories, technical package names, and Git commit histories remain 100% autonomous.

Key Highlights & Milestones Delivered

1. Dual-Backbone Architecture

Every profile is grounded by two core backbones:

  • HeLa Mitosis (chaining-mcp): Cellular division and orchestration backbone. Bundled Needle 3 agent runtime (plan → execute → observe, offline-first), Needle-backed sequential reasoning (sequentialthinking, agent_run), task decomposition, and 40 domain prompts. No external sequential-thinking server.
  • HeLa Genome (Project-Guardian-mcp-server): Living memory and state backbone. SQLite knowledge graph (memory.db), entity-relation tracking, cross-session context restoration, and decision memory.

2. 10-MCP Component Taxonomy

  • HeLa Membrane (filesystem-mcp-server): Sandboxed workspace filesystem operations and search.
  • HeLa Nucleus (terminal-mcp-server): Isolated command execution with RTK token optimization.
  • HeLa Ribosome (menager-mcp-server): Interactive PTY session multiplexing and Regex hook engine.
  • HeLa Enzyme (research-assistant-mcp-server): Unified Google Custom Search + Wikipedia caching.
  • HeLa Cytosol (Browser-Agent): Playwright browser automation and DOM accessibility perception.
  • HeLa Phenotype (the-designer): UI/UX design tokens, OKLCH palettes, and Tailwind component synthesis.
  • HeLa Receptor (scrcpy-mcp): Android mobile automation via ADB bridge.
  • HeLa Plastid (ll3m-agent): Autonomous Blender procedural 3D modeling and rendering.

3. Diagnostic Health Checker (setup.sh doctor)

Comprehensive automated diagnostic tool auditing host prerequisites (Node.js >=18, Git, SQLite3), server builds, commit pins, stdio JSON-RPC smoke readiness (<300ms), external runtimes (Playwright, ADB, Blender), and zero-key offline fallback modes.

4. 70-Combination Multi-Client Matrix (npm run test:matrix)

Full verification across all 10 client backends (Cursor, Claude, Gemini, Antigravity, OpenCode, Kilo, Zed, Codex TOML, Docker Compose, Skip) and 7 agent profiles with 100% schema accuracy.

5. Master Integration Test Suite (npm test)

Automated stdio JSON-RPC handshake and tools discovery across all 10 servers (295 active tools discovered and validated in 2.5s).

6. Interactive Documentation Website (docs/site/)

Modern, responsive documentation landing site featuring:

  • Interactive live Client & Profile Configurator.
  • Video showcases & embedded YouTube demos.
  • Uploaded Asciinema terminal casts.
  • Searchable 10-MCP Cellular Taxonomy directory.

Quick Start

git clone https://github.com/Hela-Ecosystem/hela-mcp-ecosystem.git
cd hela-mcp-ecosystem
./setup.sh doctor
./setup.sh

Documentation

License & Attribution

Distributed under the MIT License. Developed by Azzar.

The HeLa project respectfully recognizes Henrietta Lacks and the enduring scientific legacy of HeLa cells.