sagrada-mcp

sagrada-mcp

MCP server for tracking and verifying AI reasoning state, with signed action receipts, repo-history records, and preflight gating for tool calls.

Category
访问服务器

README

sagrada-mcp

Model Context Protocol (MCP) server for Sagrada — track and verify reasoning state for AI systems, and read attested repo-history records.

Sagrada gives Claude Code (and any MCP-capable client) durable, content-addressed memory of design decisions, claim revisions, contradictions, and the cryptographically verifiable chain of edits that produced the current knowledge state.

What this repository is

The source of the sagrada-mcp package on PyPI, so you can read what you installed instead of taking our word for it.

The repo was created after 0.3.0 shipped. Its first commit is that release's source, verbatim, tagged v0.3.0 — so the code you install and the code you read can be checked against each other:

pip download sagrada-mcp==0.3.0 --no-binary :all: --no-deps -d /tmp/pypi
tar xzf /tmp/pypi/sagrada_mcp-0.3.0.tar.gz -C /tmp
mkdir -p /tmp/tagged && git archive v0.3.0 | tar x -C /tmp/tagged
diff -r /tmp/tagged /tmp/sagrada_mcp-0.3.0 --exclude PKG-INFO   # no output

CI runs that diff on every push. main is ahead of v0.3.0 and carries unreleased work; pip install sagrada-mcp gives you the tag, not the branch.

What is not here: the compiled sagrada engine (Rust), which the 12 reasoning-tracker tools need. It installs from PyPI as sagrada-engine, and its source is not published. The preflight gate — the part that refuses — is in this repo, is pure Python, and does not depend on the engine.

Installation

Mode 1: pip install (clean venv, no repo checkout)

pip install sagrada-mcp

This installs the server plus vendored copies of its not-yet-published dependencies (see _vendor/README.md). One server, 21 tools:

  • The 3 preflight-gate tools (preflight_action, list_beliefs, verify_receipt) gate consequential tool calls against the repo's typed beliefs and emit Ed25519-signed action receipts to .sagrada/receipts.jsonl, each independently verifiable offline with uvx er1-verify (1.0.1 or later — an unpinned install against a version constraint is a declared coverage gap the verifier recomputes, a receipt shape 1.0.0 predated). Deterministic and engine-free — refusal does not depend on the binary wheel.

    Receipts are signed with a keypair created for your repo on first use: the private half lives at .sagrada/operator_key (mode 0600 on POSIX, gitignored, never leaving your machine), the public half at .sagrada/operator_key.pub is meant to be committed. verify_receipt checks a receipt's signer against that committed key and the signed succession chain behind it, so a receipt from a key that was never this repo's comes back foreign rather than passing quietly. Earlier versions minted a key per process, which made every signature unpinnable — "signed by X" where X existed for one run.

    What that establishes is continuity, not identity: if you pinned this repo's fingerprint out of band, and the private key has never left the machine, then a swap of the signing key shows up. Each clause matters — a signature proves possession of a file, and files copy.

    For Claude Code there is also an ENFORCED layer: the sagrada-preflight-hook console script can be wired as a PreToolUse hook, and then blocks HALTed calls in-band. Installing the package does not register it — nothing here edits your Claude settings. You add it yourself; preflight_hook.py carries the exact one-line entry to paste.

  • The 6 repo-record (crux) tools are fully functional given a .crux artifact (SAGRADA_CRUX). Without one they return a helpful error string.

  • The 12 reasoning-tracker tools additionally need the compiled sagrada-engine wheel (a hard dependency as of 0.3.0; installed automatically on platforms with wheels). If the engine is somehow absent they return a clear JSON error naming the missing piece — never a traceback.

Mode 2: from a clone of this repo

git clone https://github.com/Cruxia-Labs/sagrada-mcp && cd sagrada-mcp
pip install -e ".[dev]"
pytest -q

To run the server itself, note that it speaks MCP over stdio and does not exit — it waits for a client, so run it from your MCP client's config rather than expecting a prompt back:

python -m sagrada_mcp.server     # blocks; Ctrl-C to stop

The server prefers the real sagrada_tracker / sagrada_preflight packages when they are importable and falls back to the copies in sagrada_mcp/_vendor/ otherwise — so it behaves identically here and in the upstream tree those copies come from. In a clone the vendored copies are what runs, and the test suite binds the same ones the server did (tests/conftest.py prints which, in the pytest header).

Configure in Claude Code / Claude Desktop

{
  "mcpServers": {
    "sagrada": {
      "command": "sagrada-mcp",
      "env": { "SAGRADA_PROJECT": "/path/to/your/project" }
    }
  }
}

(In a repo checkout use "command": "python", "args": ["-m", "sagrada_mcp.server"].)

Tools provided (21)

Preflight gate (3) — deterministic, no engine required

Tool Purpose
preflight_action Check a proposed tool call against the repo's typed beliefs; returns ALLOW or HALT with a signed receipt
list_beliefs What the gate currently enforces, and which beliefs are HALT-eligible
verify_receipt Re-verify a receipt: signature, recomputed verdict, and whether the signer is this repo's operator

Recognition is a finite, best-effort vocabulary. The gate maps a tool call to beliefs by reading the forms an agent actually emits — imports, package-manager installs, deploy targets, env assignments. An action expressed in a form it does not recognise is not mapped, and an unmapped action does not conflict with anything, so it is ALLOWed. Treat HALT as an enforced policy check on recognised actions, not as containment.

Reasoning tracker (12) — require the compiled sagrada engine

Tool Purpose
track_decision Track a single design decision (auto-classifies as assert / refine / revise)
track_from_text Extract and track decisions from free text
ingest_text Ingest free text — auto-extracts claims, flags contradictions per claim
query_decision Look up a tracked decision by term
search_decisions Search decisions by term or definition
get_decision_history Evolution history of a decision (or all)
get_reasoning_status Full reasoning state summary
check_coherence Report contradictions and coherence score
check_consistency Read-only check whether new text contradicts existing knowledge (needs the optional ML battery for NLI)
verify_chain Verify cryptographic integrity of the reasoning chain
retract_decision Retract a previously tracked decision
what_changed Prose summary of tracker-state changes since a record index (not a diff: each bucket is truncated to 5 names)

Repo-record / crux (6) — work anywhere, given a .crux artifact

Deterministic reads of a transformed repo-history record — no model, no network, no clock. Every answer carries the record reference (artifact, state_root, receipt) for offline verification.

Tool Purpose
repo_asof What value a tracked repo key held on a given date (with held-from/until evidence)
what_replaced What value replaced a key's previous value, and when (flags retractions and comebacks)
stale_rule_check List a tracked file's retired values and died-and-came-back values, with evidence
crux_status The record being read: artifact, source HEAD, counts, state_root, receipt
key_history Every recorded value of one tracked key, oldest first, with the record reference
key_status Whether a tracked key is live, retired, or died-and-came-back

There is no public builder for .crux artifacts yet. The reader is here and is stdlib-only, so these six tools work fully against an artifact you already have — but the tool that builds one is not in this repo and is not published. If you have no artifact, each crux tool returns a pointer rather than failing, and until a builder ships that pointer is the honest end of the road for a public user.

Resources provided (3)

Resource URI Content
sagrada://state/summary Concise state summary for LLM context
sagrada://state/decisions All tracked decisions
sagrada://state/contradictions Currently-detected contradictions

Environment variables

Variable Purpose
SAGRADA_PROJECT Path to the Sagrada project directory (auto-init if missing)
SAGRADA_NO_ML Set to true to disable the ML classification battery (deterministic Jaccard floor instead)
SAGRADA_BACKEND Persistence backend: json_file (default) or sqlite
SAGRADA_CRUX Path to the .crux artifact the repo-record tools read (falls back to results/*.crux under SAGRADA_PROJECT)

License

Apache-2.0. Published by Cruxia-Labs.

推荐服务器

Baidu Map

Baidu Map

百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。

官方
精选
JavaScript
Playwright MCP Server

Playwright MCP Server

一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。

官方
精选
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。

官方
精选
本地
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。

官方
精选
本地
TypeScript
VeyraX

VeyraX

一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。

官方
精选
本地
graphlit-mcp-server

graphlit-mcp-server

模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。

官方
精选
TypeScript
Kagi MCP Server

Kagi MCP Server

一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。

官方
精选
Python
e2b-mcp-server

e2b-mcp-server

使用 MCP 通过 e2b 运行代码。

官方
精选
Neon MCP Server

Neon MCP Server

用于与 Neon 管理 API 和数据库交互的 MCP 服务器

官方
精选
Exa MCP Server

Exa MCP Server

模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。

官方
精选