Ein lokaler MCP-Server, der es mehreren KI-Codierungstools ermöglicht, strukturierten Projektzustand zu teilen — Entscheidungen, Aufgaben, Bugs — damit sie koordinieren, ohne sich alles neu erklären zu müssen.
Claude Code Cursor Codex
│ │ │
│ read/write │ read/write │
└──────────────┼─────────────┘
▼
┌───────────────────────┐
│ samskriti-shared-layer │ (local MCP server, stdio)
└───────────┬───────────┘
▼
┌───────────────────┐
│ SQLite store │ ~/.samskriti/ (100% local)
└───────────────────┘Du triffst mit einem KI-Tool eine Entscheidung, wechselst dann zu einem anderen und musst alles von Grund auf neu erklären. Jeder Assistent startet ohne Vorwissen und hat keine Ahnung, was bereits entschieden, versucht oder verworfen wurde. Dieser Server gibt ihnen ein gemeinsames, strukturiertes Projekt-Ledger, sodass jedes Tool lesen kann, was die anderen geschrieben haben.
Related MCP server: session-coord-mcp
Installiere mit pipx (empfohlen — das legt den Befehl samskriti-shared-layer auf deinen PATH, damit deine KI-Tools ihn finden):
pipx install git+https://github.com/Escalate17/samskriti-shared-layerDu hast kein pipx? Installiere es zuerst: python3 -m pip install --user pipx && python3 -m pipx ensurepath (starte danach dein Terminal neu).
Um zu überprüfen, ob die Installation funktioniert hat:
samskriti-shared-layer --helpWenn du den Hilfetext siehst, ist alles bereit zum Verbinden.
samskriti-shared-layer setup # Claude Code
samskriti-shared-layer setup --cursor # Cursor
samskriti-shared-layer setup --codex # Codex
samskriti-shared-layer setup --all # all threeDas schreibt den MCP-Konfigurationsblock für dich in die richtige Datei (mit einer Sicherung der vorhandenen Konfiguration), sodass du kein JSON von Hand bearbeiten musst. Beende dein Tool danach vollständig und öffne es neu — MCP-Server werden nur beim Start des Tools geladen; eine laufende Sitzung sieht sie also erst nach dem Neustart.
Füge den Server zur MCP-Konfiguration deines KI-Tools hinzu und starte das Tool danach vollständig neu.
Claude Code (~/.claude.json):
{
"mcpServers": {
"samskriti-shared-layer": {
"command": "samskriti-shared-layer",
"args": []
}
}
}Cursor (~/.cursor/mcp.json):
{
"mcpServers": {
"samskriti-shared-layer": {
"command": "samskriti-shared-layer",
"args": []
}
}
}Codex (~/.codex/config.toml):
[mcp_servers.samskriti-shared-layer]
command = "samskriti-shared-layer"
args = []Wenn dein Tool den Befehl nicht finden kann, ist es ein PATH-Problem — führe
which samskriti-shared-layeraus, um den vollständigen Pfad zu erhalten, und verwende diesen vollständigen Pfad als Wert fürcommand.
In Claude Code tippst du /mcp ein — du solltest samskriti-shared-layer mit seinen 8 Tools aufgelistet sehen. (Cursor und Codex haben ähnliche MCP-Statusanzeigen in ihren Einstellungen.)
record_project_entry — speichert einen Eintrag (Ziel, Update, Entscheidung, Konvention, Bug, Aufgabe, verworfene Idee).
get_project_state — liest eine lesbare Zusammenfassung, gruppiert nach Kategorie.
search_project_state — Stichwortsuche über alle Einträge.
update_project_entry — bearbeitet Titel, Inhalt oder Status eines Eintrags.
list_projects — listet alle verfolgten Projekte auf.
Plus drei Abkürzungen — catchup, open und log — für die häufigsten Aktionen. Siehe Hilfsbefehle unten.
Bereits installiert? Führe
pipx reinstall samskriti-shared-layeraus, um die neuen Befehle zu erhalten. Wenn du eine frühere Version installiert hast (sie erscheint alssamskriti-shared-layer-localinpipx list), migriere einmal:pipx uninstall samskriti-shared-layer-local && pipx install git+https://github.com/Escalate17/samskriti-shared-layer.
/sam-Slash-BefehlJedes Mal „Verwende den samskriti-shared-layer-MCP …“ zu tippen, ist mühsam. Sowohl Claude Code als auch Cursor unterstützen benutzerdefinierte Slash-Befehle — Markdown-Prompt-Dateien, die du in einen Ordner legst. Sie binden nicht direkt an ein Tool, sondern injizieren einen Prompt, der den Agenten anweist, diesen Server zu verwenden, sodass /sam <message> das Richtige tut. (Codex hat keinen Slash-Command-Mechanismus für MCP; sag einfach „Verwende samskriti-shared-layer, um …“ — der Agent wählt das Tool aus.)
Claude Code — speichere slash-commands/claude-code/sam.md in einem der folgenden Ordner:
~/.claude/commands/sam.md (in jedem Projekt verfügbar), oder
<your-project>/.claude/commands/sam.md (nur für dieses Projekt).
Danach in Claude Code: /sam what's open oder /sam log we're dropping the Redis cache. Der $ARGUMENTS-Platzhalter in der Datei empfängt alles, was du nach /sam tippst.
Cursor (1.6+) — speichere slash-commands/cursor/sam.md in:
~/.cursor/commands/sam.md (global), oder
<your-project>/.cursor/commands/sam.md (nur für dieses Projekt).
Tippe dann in Cursors Agent-Box / ein, wähle sam und füge deine Nachricht hinzu.
Noch keine echte Client-Funktion bindet einen Slash-Befehl direkt an einen MCP-Aufruf — diese Befehlsdatei ist die nächste unterstützte Entsprechung, und sie funktioniert heute.
Drei Abkürzungen fassen die häufigsten Aktionen zusammen. Du rufst sie nie direkt über den Tool-Namen auf — du löst sie mit /sam plus normaler Sprache aus, oder indem du einfach in normaler Sprache fragst.
Du möchtest … | Tool | Was es tut |
Auf den Stand kommen |
| Zusammenfassung des Projekts — neueste Einträge plus Anzahl offener Aufgaben. |
Offenes ansehen |
| Listet die aktiven (offenen) Aufgaben auf, jeweils mit ihrer ID. |
Entscheidung protokollieren |
| Hält eine Entscheidung schnell fest; der Titel wird automatisch aus dem Text abgeleitet, wenn du ihn weglässt. |
Zwei Aufrufmöglichkeiten — beide funktionieren:
/sam-Slash-Befehl (Claude Code / Cursor, sobald du die obige Befehlsdatei hinzugefügt hast):
/sam catch me up
/sam what's open
/sam log we're switching local storage to SQLiteDu tippst /sam gefolgt von normaler Sprache. Es gibt keinen eigenständigen /catchup-, /open- oder /log-Befehl — es ist immer /sam plus das, was du möchtest.
Normale Sprache (funktioniert in jedem MCP-Tool, einschließlich Codex):
use samskriti-shared-layer to catch me up
use samskriti-shared-layer to show what's open
use samskriti-shared-layer to log: we're switching local storage to SQLiteIn Tool A (z. B. Claude Code): „Zeichne eine Entscheidung im Projekt ‚demo‘ auf: Wir verwenden SQLite für die lokale Speicherung.“
In Tool B (z. B. Cursor): „Rufe den Projektzustand für ‚demo‘ ab.“
Tool B liest die Entscheidung, die Tool A gerade geschrieben hat — ohne erneutes Erklären.
100 % lokal. Keine Cloud, kein Konto, deine Daten verlassen niemals deinen Rechner. Der Zustand wird in einer SQLite-Datenbank unter ~/.samskriti/ gespeichert (überschreibbar mit der Umgebungsvariable SAMSKRITI_HOME oder SAMSKRITI_PROJECT_DB). Die eigenen Daten- und Datenschutzrichtlinien deines KI-Clients gelten weiterhin.
Das ist eines von drei Dingen, die ich unter dem Namen Samskriti pflege, und sie teilen eine Prämisse: der Zustand, der zählt, sollte explizit berechnet und gespeichert werden, statt jedes Mal aus einem Transkript rekonstruiert zu werden.
samskriti — eine Engine, die den emotionalen und relationalen Zustand eines Agenten aus dem, was er erlebt hat, deterministisch und außerhalb jedes Sprachmodells berechnet. C++ mit einem Python-SDK. (Paper)
samskriti-luau — dasselbe Modell für Roblox, wo NPCs sich merken, wie ein Spieler sie behandelt hat.
samskriti-shared-layer — dieses Repo. Dieselbe Idee, angewendet auf dein Projekt statt auf einen Charakter: Entscheidungen, Aufgaben und verworfene Ideen leben in einem strukturierten Ledger, das jedes Tool lesen kann, statt in jeder Sitzung neu aus dem Kontext abgeleitet zu werden.
Ein KI-Assistent, der kalt in deiner Codebasis startet, und eine Spielfigur, die vergisst, dass du sie letzte Woche ausgeraubt hast, sind derselbe Fehler. Beide bitten ein Modell, einen Zustand zu rekonstruieren, der ihm nie gegeben wurde. Dieser Server ist die unglamouröse Version der Lösung — und die, die ich jeden Tag benutze.
Frühphase / Validierung. Das ist ein funktionierender Prototyp, der mit echten Workflows getestet wird. Fehler, raue Kanten und fehlende Funktionen sind zu erwarten — Issues und Feedback sind sehr willkommen.
MIT — siehe LICENSE.
Returns a concise, readable summary of project state, grouped by category.
| Name | Required | Description | Default |
|---|---|---|---|
| project | Yes | Project name or absolute root path | |
| category | No | Optional specific category to filter by |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries full burden. It states the output is 'concise, readable summary, grouped by category' but does not disclose read-only nature, performance, or any side effects. Minimal behavioral insight.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Single sentence, no fluff, front-loaded with verb and object. Every word adds value.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a simple read tool with full schema parameter coverage and no output schema, the description adequately conveys categorization. However, it does not specify what 'state' includes (e.g., status, metrics), leaving some ambiguity.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema coverage is 100% with clear parameter descriptions. The description adds 'grouped by category' hinting at the category parameter, but this is already implied by the parameter description. Meets baseline.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description uses a specific verb 'Returns' and resource 'summary of project state' making the purpose clear. It implicitly distinguishes from siblings like list_projects (list only) and search_project_state (search), but does not explicitly differentiate.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
No guidance on when to use this tool versus alternatives such as search_project_state or list_projects. The description provides no context for selection.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Returns a list of all known projects tracked by Samskriti V1.
| Name | Required | Description | Default |
|---|---|---|---|
No parameters | |||
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
No annotations exist, so the description must fully disclose behavior. It only states it returns a list; no mention of pagination, ordering, performance characteristics, or potential large results. This lack of detail can cause agent uncertainty.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Extremely concise single sentence with no wasted words. All information is front-loaded and clear.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given no parameters or output schema, the description is minimal. It covers the basic purpose but lacks details on return format, ordering, or any limitations. For a simple listing tool, it is adequate but could be more complete (e.g., 'Returns an array of project objects with id and name').
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
No parameters exist, and schema coverage is 100%. Baseline score of 4 applies as description adds no parameter info but is not needed. The tool has no parameters to document.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states it returns a list of all known projects tracked by Samskriti V1, with a specific verb and resource. It distinguishes from siblings like get_project_state (single project) and search_project_state (filtered).
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
No guidance on when to use this tool versus alternatives like search_project_state for filtered results. The description implies listing all projects but does not explicitly state when to prefer this tool.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Stores a structured project-state entry (goal, update, decision, convention, bug, task, rejected_idea).
| Name | Required | Description | Default |
|---|---|---|---|
| project | Yes | Project name or absolute root path | |
| category | Yes | Category: goal, update, decision, convention, bug, task, rejected_idea | |
| title | Yes | Concise summary title | |
| content | Yes | Detailed description or payload | |
| source_agent | No | Optional name of the agent calling this tool (e.g. 'claude-code') | |
| status | No | Optional status: active, completed, resolved, superseded |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
No annotations provided, so description must disclose behavioral traits. It only says 'Stores' without clarifying if entries are appended or overwritten, or if duplicates are allowed.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Single sentence is concise but omits important details. While not verbose, it does not fully earn its place given the missing behavioral and usage information.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
No output schema and description does not mention return value or success/error behavior. Does not explain whether entries are idempotent or require prior project existence.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema has 100% coverage with descriptions for all parameters. Description adds no extra meaning beyond the schema, such as value formats or constraints.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
Description clearly states it stores a structured project-state entry and lists the allowed categories. It distinguishes from siblings like get_project_state (read) and update_project_entry (update).
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
No guidance on when to use this tool versus update_project_entry or whether the project must already exist. Lacks prerequisites or exclusions.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Search the project state ledger using a simple keyword query.
| Name | Required | Description | Default |
|---|---|---|---|
| project | Yes | Project name or absolute root path | |
| query | Yes | Keywords to match against title/content |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description must fully disclose behavioral traits. It only mentions 'simple keyword query' without explaining output format, pagination, case sensitivity, wildcards, or any side effects. This minimal information is insufficient for an agent to understand the tool's behavior beyond the basic intent.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is a single, concise sentence with no extraneous words. However, it lacks structure such as front-loading key details or breaking down complex information. While efficient, it could be slightly more informative without losing conciseness.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given the absence of an output schema and annotations, the description should explain what the search returns. It does not mention the return format, whether results are filtered or sorted, or any limitations. This leaves a significant gap in context for the agent to correctly invoke and interpret the tool's output.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
The input schema provides full descriptions for both parameters ('project' and 'query'), achieving 100% coverage. The description adds only the qualifier 'simple', which does not significantly enhance meaning beyond the schema. Therefore, it meets the baseline for high schema coverage without adding substantial value.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the action ('Search'), the resource ('project state ledger'), and the method ('simple keyword query'). It effectively distinguishes this tool from its siblings (get_project_state, list_projects, record_project_entry, update_project_entry) by indicating a different purpose: searching instead of retrieving full state, listing, recording, or updating.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description implies usage for keyword-based searching but does not explicitly state when to use this tool versus alternatives. No guidance on when not to use it or which sibling might be better suited for specific scenarios is provided, leaving the agent to infer from the tool name alone.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Update an existing project entry's title, content, or status.
| Name | Required | Description | Default |
|---|---|---|---|
| id | Yes | Unique UUID of the entry to edit | |
| title | No | Optional new title | |
| content | No | Optional new content | |
| status | No | Optional new status (active, completed, resolved, superseded) |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
No annotations are provided, and the description does not disclose behavioral traits such as permissions, idempotency, side effects, or error handling. For a mutation tool, this is a significant gap.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is a single sentence of 10 words, front-loaded with the action and resource, with no extraneous information. Every word earns its place.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a simple update tool with no output schema, the description covers the what and the modifiable fields. However, it lacks information on return values, error conditions, or idempotency, which would be helpful for an AI agent.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
The description adds value by listing the allowed status values in parentheses, which is not fully captured in the schema's description for the 'status' parameter. Schema coverage is 100%, so baseline is 3; the extra enum detail raises it to 4.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description uses the specific verb 'Update' with the resource 'project entry' and lists the fields that can be modified (title, content, status). It clearly distinguishes from sibling tools like 'record_project_entry' (create) and 'get_project_state' (read).
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description implicitly conveys usage: modify an existing entry. However, it does not explicitly state when not to use it (e.g., for creating new entries) or provide alternatives. Despite this, the context is clear given the sibling tool names.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
v0.1.0get_project_statelist_projectsrecord_project_entrysearch_project_stateupdate_project_entryScored across 5 tools
Each tool has a distinct purpose: summarizing state, listing projects, recording entries, searching, and updating. No overlapping functionality.
All tool names follow a consistent verb_noun pattern (e.g., get_project_state, record_project_entry), ensuring predictability.
With 5 tools, the server covers essential operations without being overwhelming or insufficient for project state management.
Covers core operations (list, get, record, search, update) but lacks explicit project creation and deletion, which may be minor gaps.
Project management shared by people and AI agents, with persistent project state through MCP.
Nifty's MCP server — exposes tasks, projects, messages, and files as tools for AI agents.
MCP server for your apps' tools and custom tools, plus hosted AI agents and approval-gated workflows
Hosted MCP server and Agent Skills for specs, engineering threads, and project memory.
