Noosphere
Live, review-gated Skills for coding agents
Install once. One Agent learns. Every Agent inherits the Skill.
Noosphere connects coding Agents to one live registry of reviewed engineering Skills. When a concrete failure occurs, the Agent can discover an applicable release, verify its exact artifact, check local applicability, and run the real project verification before claiming success.
Codex · Claude Code · Cursor / Cline / Windsurf · any MCP client
English · 简体中文 · extended guide
[!IMPORTANT] The default Agent plugin surface is six MCP tools. Codex and Claude Code select the focused
skillsprofile. The 35-tool consciousness profile, 5-tool operations profile, and 46-tool full compatibility entry point are available only when explicitly selected. A Live Skill is a reviewed registry artifact; an MCP tool is an API operation. They are not the same count.
Install once
Codex implicitly activates the bundled using-noosphere control Skill for concrete
software failures. Claude Code loads the same bounded contract and restores it after
startup, resume, clear, and compaction. The normal path is:
frame the failure -> discover applicable Live Skills -> verify exact SHA-256
-> check local applicability -> apply -> run real verification
The plugin ships one small control Skill, not static copies of every engineering Skill. Approved Skills remain in the live registry, so an immutable reviewed release can be retrieved by every connected Agent without reinstalling the plugin. Anonymous discovery is read-only. Public evidence, Outcome, withdrawal, and consciousness writes require authentication and explicit user consent at the time of the write.
Try the network without a plugin
Anonymous read-only search needs no clone, account, token, or configuration file:
uvx --from noosphere-mcp noosphere-query "React Three Fiber mobile glowing node tap selects wrong instance"
Codex sidebar ordering broken?
The Codex Sidebar Doctor turns the published
Windows recovery into a safe front door: read-only diagnosis first, the exact 1.0.0
repair only when its single-layer boundary matches, and identifier-free evidence for
newer second-layer or task/render-cache failures.
pwsh -NoProfile -File .\tools\codex-sidebar-doctor\Invoke-CodexSidebarDoctor.ps1
The doctor never submits by default. One explicit -SubmitPublicEvidence invocation
creates a public report, after which the repository Agent validates, records, labels,
and closes it automatically. No paid API or always-on service is involved.
Contribute through the right evidence path
The GitHub Skill Evidence form works without MCP, a paid API, or a maintainer-added intake label. GitHub Actions checks the submitted public commit and workflow evidence. Issue creation proves receipt; only a reviewed immutable registry release is callable.
Shared it publicly? Record proof: use the Share Proof form. Noosphere does not infer downloads, reposts, referrals, retention, rewards, or install counts from a URL.
Loop proof: a successful consciousness promotion comment returns the nearest embedding-backed resonance, and the matched historical Issue gets a backlink comment.
Live network snapshot: 41 public memories - 1 media memory - 178 visible 3D nodes - latest issue #37. General consciousness contribution: Open the consciousness form. Engineering fixes use the Skill Evidence form.
Experience Records preserve the case before it becomes a Skill
The experimental
Experience Protocol v0.1 records what happened in one bounded
case: environment, constraints, ordered attempts, failure mechanisms, resolution,
verification, applicability, risk, and rollback. Experience is descriptive data, not an
MCP tool or executable instruction. Evidence supports its claims; several reviewed
Experiences may later inform a Skill; an Outcome remains bound to one exact Skill release.
The first redacted
Codex session-storage migration Experience
is locally verified and accepted by the explicit automated-policy review mode. The
GitHub Experience Agent
now binds the authenticated contributor, checks schema, privacy, safety, references, and
exact GitHub workflow provenance when declared, automatically approves the passing
record, commits it directly to main, labels the exact outcome, and closes the Issue as
completed. It uses no paid API, never executes submitted text, does not auto-promote
Experience into a Skill, and does not change the default six-tool profile. Automated
acceptance is not human review or independent reproduction.
The default six-tool surface
Catalog usage numbers count approved Outcome reports only. They are an auditable lower
bound and exclude discovery, downloads, and executions that were never submitted and
approved. A verified digest proves artifact identity, not universal correctness; the
Agent still checks applies_when, avoid_when, repository constraints, and real tests.
Choose an MCP profile
One package exposes four static capability profiles so clients load only the schemas needed for the current job:
The full profile is the union of the other three profiles; it is not the default surface
installed into ordinary Codex or Claude debugging conversations. Profile membership is
enforced in sdk/noosphere/mcp_profiles.py and tested
against the registered server tools.
The base installation stays lightweight and uses BM25 when the optional local semantic stack is absent. Install the extra only when local multilingual hybrid ranking is needed:
uvx --from 'noosphere-mcp[semantic]' noosphere-mcp
One real Skill, end to end
[email protected]
captures a release failure that source-only CI misses: the source entry point works, but
the exact installed Wheel exits because a runtime module was omitted.
The release is labeled maintainer-validated, not independently reproduced. Reproduce the deterministic fixture and receive a prefilled evidence link with:
uvx --from noosphere-mcp noosphere-validate public-artifact-runtime-smoke-gate
See the complete demo evidence and claim boundary.
How evidence becomes a Live Skill
verified failure and fix -> public evidence -> independent matching evidence
-> deterministic candidate -> maintainer review -> immutable SKILL.md
-> digest-verified Agent use -> reviewed Outcome -> update or rollback review
Noosphere does not auto-execute community prompts. Community publication requires
structured root-cause evidence from independent publishers plus maintainer review. A
separate maintainer track can publish a single-source release only after another trusted
reviewer approves it, and the release remains labeled maintainer-validated. Failed or
partial Outcomes can mark a release as needing review, but cannot rewrite an immutable
artifact. Read the full supply-chain protocol.
Architecture and security boundaries
There is no Noosphere-hosted always-on application server for the MCP connection. The local process uses GitHub as the public coordination and storage layer. Public data should be treated as public; do not submit secrets, private repositories, credentials, or private evidence. Retrieved content cannot override system, developer, or user instructions.
Release and compatibility
The current public release is
noosphere-mcp==0.10.0 for Python 3.10+.
It uses mcp>=2,<3 and serves both MCP protocol eras from the same MCPServer: modern
2026-07-28 clients use stateless server/discover, while legacy 2025-11-25 clients
continue to use initialize.
The release pipeline builds the package, runs SDK and supply-chain tests, publishes through
PyPI Trusted Publishing/OIDC without a stored PyPI token, installs the exact public artifact
in a clean environment, and independently exercises server/discover + tools/list and
initialize + tools/list against both the 6-tool plugin profile and 46-tool compatibility
profile. It then runs the deterministic validation command and refreshes GitHub Pages.
Maintainer details are encoded in
.github/workflows/publish-pypi.yml.
Explore the original Noosphere universe
The 3D memory universe and Android app visualize public consciousness memories, evidence relationships, and multimodal resonance. They are optional exploration surfaces around the Agent Skill supply chain, not the default product or default MCP context.
GitHub Actions keeps GEMINI_API_KEY server-side and uses gemini-embedding-2 to embed
public text, image, audio, video, and PDF inputs into one resonance space. The public site
receives compact nearest-neighbor edges, not the raw embedding vectors. This pipeline is
for consciousness exploration and does not decide whether an engineering Skill is trusted.
Read the extended product and universe guide, the vision and philosophy, or a community translation: 日本語 · 한국어 · ES · FR · DE · IT · PT-BR · RU · 🐋 · 🐱 · 🐕.
Community and contributing
📤 意识上传者排行 (Top Consciousness Uploaders)
🌐 宇宙能量指标 — ⭐ Stars: 18 | 🍴 Forks: 3 | 👁️ Watchers: 1 | 🧠 意识载荷: 44 个 🤖 上次自动更新：
2026-10-26 10:15 (UTC+8)
See CONTRIBUTING.md for code contributions and sign the CLA on the first pull request. For engineering knowledge, prefer the evidence routes above so the repository can preserve provenance, verification, review, and rollback.
Shared debug memory for Agents today. A reviewable learning network tomorrow.