Skip to main content
Glama
buildoak

eywa-mcp

by buildoak
OverviewSchemaSourceRelated ServersDiscussionsAdmin
Python
Local

Eywa MCP

Cross-session memory for Claude Code (MCP server + CLI).

Build PyPI License

The Problem

Claude Code sessions are ephemeral. Context is lost between sessions, so you start fresh each time and re-explain what you were working on.

For heavy Claude Code users with hundreds of sessions, that context reset becomes a major productivity drain.

Related MCP server: agent-mem0

The Solution

Eywa extracts a structured handoff at the end of each session and retrieves relevant past context at the start of the next one.

The name comes from the neural network in Avatar: Eywa connects sessions the way Eywa connects living memory.

How It Works

Eywa runs a deterministic pipeline around your Claude Code transcripts:

  1. Session Detection: 4-strategy fallback (explicit session ID, PID tracing, CWD mtime, global mtime).

  2. Session Conversion: JSONL transcript -> normalized markdown conversation.

  3. Extraction: LLM-powered structured handoff extraction.

  4. Indexing: Inverted index with metadata + TF-IDF-friendly keyword/project maps.

  5. Retrieval: Query keyword scoring + recency decay to return relevant handoffs.

Claude Code JSONL Session
          |
          v
 [Session Detection]
          |
          v
 [JSONL -> Markdown]
          |
          v
 [Structured Extraction]
          |
          v
 [Handoff Markdown + Index]
          |
          v
      eywa_get()

Two-Stage Setup

Stage 1: Batch Index (one-time setup)

Use eywa-batch to process existing historical sessions in bulk through OpenRouter.

  • You can pick any OpenRouter model (Gemini Flash, Claude, GPT, Llama, etc.)

  • Default batch model: google/gemini-3-flash-preview

  • Designed for hundreds of prior sessions

  • Fast + low-cost extraction pass

  • Builds your initial handoff corpus and index

Stage 2: Runtime

Run eywa-mcp alongside Claude Code for ongoing sessions.

  • Uses Claude (Sonnet) extraction at session end (eywa_extract())

  • Retrieves relevant context at session start (eywa_get())

  • Installs a companion CLI (eywa) for scripts and manual use

Installation

Prerequisites

  • Python 3.10+

  • Node.js 18+

  • Claude Code

Option A: Bootstrap (recommended)

Run the repo bootstrap script to check prerequisites and install both Python and Node dependencies:

./setup.sh

This installs three commands:

  • eywa-mcp (MCP stdio server)

  • eywa (CLI: get/extract/rebuild-index)

  • eywa-batch (OpenRouter-powered batch indexing)

Option B: Manual install

1) Install Python package (editable)

pip install -e .

2) Install Node extractor dependencies

cd eywa/extractors
npm install
cd ../..

3) Configure environment

cp .env.example .env

4) Register MCP server

Add to claude_desktop_config.json or ~/.claude.json:

{
  "mcpServers": {
    "eywa": {
      "command": "eywa-mcp"
    }
  }
}

5) Run manually (optional)

eywa-mcp

Configuration

Variable

Default

Description

EYWA_DATA_DIR

~/.eywa

Runtime storage root for handoffs and index

EYWA_SESSIONS_DIR

~/.claude/projects

Claude Code session JSONL root

EYWA_TASKS_DIR

<EYWA_SESSIONS_DIR parent>/tasks

Tasks directory used for PID-based session detection

EYWA_CLAUDE_MODEL

sonnet

Model used by runtime extraction (eywa_extract)

EYWA_OPENROUTER_MODEL

google/gemini-3-flash-preview

OpenRouter model used by batch indexing (eywa-batch)

OPENROUTER_API_KEY

(unset)

OpenRouter API key for batch extraction

EYWA_BATCH_DELAY

0.5

Delay (seconds) between batch API calls

EYWA_BATCH_CONCURRENCY

5

Concurrent sessions processed by eywa-batch

EYWA_TIMEZONE

UTC

Timezone for rendered session timestamps

EYWA_LOG_LEVEL

INFO

Logging verbosity

Usage

Eywa exposes two MCP tools (for Claude Code) and a CLI (for humans/scripts).

eywa_get()

Retrieve relevant context from prior handoffs.

No query (recent sessions):

{"max_handoffs": 3}

With query:

{"query": "mcp tool routing and index scoring", "days_back": 30, "max_handoffs": 4}

With tighter options:

{"query": "release pipeline", "days_back": 7, "max_handoffs": 2}

Sample output:

## Eywa: 2 past sessions

# Implemented MCP routing fallback logic

## What Happened
- Added explicit tool dispatch guard for unknown tool names.
- Introduced parse-time validation for input payload constraints.

## Open Threads
- Add integration tests for malformed tool inputs.

eywa_extract()

Extract and persist a handoff from the active session.

Auto-detect active session:

{}

Explicit session ID:

{"session_id": "12345678-1234-1234-1234-123456789abc"}

CLI (eywa)

Manual equivalents of the MCP tools:

eywa get                          # 3 most recent sessions
eywa get "mcp tool routing" --days-back 30 --max 5
eywa extract                      # auto-detect current session
eywa extract 1b2f6f6b             # 8-char short ID
eywa extract 1b2f6f6b-65a6-...    # full UUID
eywa rebuild-index                # rebuild index from stored handoffs

Batch Indexing

Run one-time bulk import of historical sessions:

eywa-batch

Set your OpenRouter API key first:

export OPENROUTER_API_KEY=...

Choose a model (optional):

export EYWA_OPENROUTER_MODEL=anthropic/claude-3.5-sonnet

Dry run (no API calls):

eywa-batch --dry-run

Custom delay between calls:

eywa-batch --delay 1.0

Set concurrency (1-20):

eywa-batch --concurrency 10

Limit the run:

eywa-batch --max 50

Force reindex all sessions:

eywa-batch --reindex

What to expect:

  • Scans EYWA_SESSIONS_DIR for *.jsonl

  • Skips already-indexed sessions (unless --reindex)

  • Skips very short/trivial sessions

  • Uses OpenRouter Chat Completions with your selected model

  • Writes handoffs to YYYY/MM/DD/<session_id>.md

  • Updates handoff-index.json incrementally

  • Prints progress and end-of-run summary

License

MIT. See LICENSE.

Available Tools

2 tools
eywa_extractA

Extract a handoff from the current session (or a specified session).

Called at end of session to persist a handoff document. Extracts key decisions, insights, and open threads via the Anthropic Claude model configured by EYWA_CLAUDE_MODEL (default: sonnet) using the Claude Agent SDK.

  • No args: auto-detects current session via PID tracing + mtime.

  • With session_id: extracts that specific session.

Examples:

  • eywa_extract()

  • eywa_extract(session_id="1b2f6f6b-65a6-42ff-aca7-34889b422799")

ParametersJSON Schema
NameRequiredDescriptionDefault
session_idNoSession UUID. Auto-detected if omitted.

TDQS

A4.5/5.0
Behavior4/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

With no annotations provided, the description bears full responsibility for behavioral disclosure. It reveals the use of the Anthropic Claude model, auto-detection via PID tracing and mtime, and persistence of a handoff document. It does not mention potential side effects or idempotency, but the disclosed behaviors are sufficient for a low-complexity tool.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness5/5

Is the description appropriately sized, front-loaded, and free of redundancy?

The description is concise (under 80 words), well-structured with a clear overview, bullet points for usage, and examples. Every sentence adds information without redundancy or fluff.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness4/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

Given the tool's low complexity (single optional parameter, no output schema), the description covers purpose, usage, parameter behavior, and underlying mechanism. It lacks explicit mention of the return value or error handling, but the intended audience can infer the output (a handoff document). It is largely complete for effective use.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters5/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

Despite high schema coverage (100%), the description adds significant value beyond the schema by explaining the auto-detection mechanism for the optional session_id, providing examples, and clarifying behavior in both arg-presence scenarios. This transforms a simple parameter into a well-understood feature.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description clearly identifies the tool as extracting a handoff from a session, with a specific verb ('Extract') and resource ('handoff from session'). It distinguishes itself from the sibling tool 'eywa_get' by focusing on extraction and persistence at session end.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines4/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

The description states the typical use case ('Called at end of session to persist a handoff document') and explains the two invocation modes (no args vs. with session_id). However, it does not explicitly exclude alternative scenarios or compare with 'eywa_get' for differentiation.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

eywa_getA

Retrieve past session handoffs for context continuity.

Called at session start or when you need context about past work.

  • No query: returns 3 most recent substantial sessions.

  • With query: keyword-matches against past sessions, returns top matches.

Examples:

  • eywa_get()

  • eywa_get(query="sorbent reasoning tokens")

  • eywa_get(query="river mcp", days_back=30, max_handoffs=5)

ParametersJSON Schema
NameRequiredDescriptionDefault
queryNoWhat are we working on? Keywords, project name, topic.
days_backNoHow far back to search (default 14 days).
max_handoffsNoHow many handoffs to return (default 3, max 5).

TDQS

A4.4/5.0
Behavior4/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

No annotations are provided, so the description bears full responsibility. It discloses that without a query it returns the 3 most recent substantial sessions, and with a query it keyword-matches and returns top matches. Default values for parameters are mentioned. It does not explicitly state it's a read-only operation, but retrieval implies non-destructive behavior. The lack of destructive hints is acceptable given the retrieve nature.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness4/5

Is the description appropriately sized, front-loaded, and free of redundancy?

The description is well-structured with a brief introductory sentence, usage notes, bullet-pointed behaviors, and examples. It is concise without unnecessary elaboration. The examples are helpful. Minor redundancy in the usage guidelines section, but overall efficient.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness5/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

Given that there is no output schema and only three optional parameters with full schema coverage, the description provides complete guidance on when and how to use the tool. It covers both query and no-query scenarios, default values, and result counts. The agent has enough information to invoke the tool correctly without external context.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters4/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

The input schema covers 100% of the parameters with descriptions. The description adds value by explaining the behavior difference with and without the query parameter, and by providing examples that illustrate how days_back and max_handoffs affect results. This supplements the schema definitions well.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description clearly states the tool retrieves past session handoffs for context continuity. The verb 'retrieve' and resource 'past session handoffs' are specific. The examples further clarify the purpose. While the sibling 'eywa_extract' exists, the description sufficiently distinguishes by focusing on handoffs retrieval.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines4/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

The description explicitly says when to use the tool: 'Called at session start or when you need context about past work.' It also explains the two main usage modes (no query vs with query) and provides examples. However, it does not explicitly state when NOT to use it or mention alternatives, but the guidance is clear enough.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

  1. 2 tool updatesv0.1.0
    • First observedeywa_extract
    • First observedeywa_get
Deploy Server
One endpoint for every MCP toolRayrun runs every MCP server behind one URL and holds the credentials. Each call is decided per client, tool, and argument, then recorded.ray.runAd

TDQS

A4.4/5.0

Scored across 2 tools

Disambiguation5/5

The two tools have clearly distinct purposes: eywa_extract saves a handoff from a session, while eywa_get retrieves past handoffs. There is no functional overlap.

Naming Consistency5/5

Both tools follow the same pattern: 'eywa_' prefix followed by a descriptive verb (extract, get). This is consistent and predictable.

Tool Count3/5

With only 2 tools, the surface is minimal. While it covers the core extract and retrieve operations for handoffs, it feels thin; additional management tools like delete or list could be expected.

Completeness4/5

The tools cover the primary operations of creating (extract) and reading (get) handoffs. Gaps include update and delete, but those may not be necessary for immutable handoffs. Minor missing features like listing all handoffs are partially addressed by querying.

Maintenance

ActivityInactive
ResponsivenessNo issues

Resources

Related MCP Connectors

Related MCP Servers