aleph-rlm 0.5.5

MCP server for recursive LLM reasoning—load context, iterate with search/code/think tools, converge on answers

Aleph

Aleph is an MCP (Model Context Protocol) server that enables AI assistants to analyze documents too large for their context window. By implementing a Recursive Language Model (RLM) approach, it allows models to search, explore, and compute over massive datasets without exhausting their token limits.

Key Capabilities

• Unlimited Context: Load files as large as your system RAM allows—gigabytes of data accessible via simple queries. The LLM never sees the raw file; it queries a Python process that holds the data in memory.
• Navigation Tools: High-performance regex search and line-based navigation.
• Compute Sandbox: Execute Python code over loaded content for parsing and analysis.
• Evidence Tracking: Automatic citation of source text for grounded answers.
• Recursive Reasoning: Spawn sub-agents to process document chunks in parallel.

How "Unlimited Context" Works

Traditional LLMs are limited by their context window (~200K tokens). Aleph sidesteps this entirely:

┌─────────────────┐     queries      ┌─────────────────────────┐
│   LLM Context   │ ───────────────► │   Python Process (RAM)  │
│   (~200K tokens)│ ◄─────────────── │   (8GB, 32GB, 64GB...)  │
│                 │   small results  │   └── your_file.txt     │
└─────────────────┘                  └─────────────────────────┘

• Python loads the entire file into RAM as a string
• The LLM queries it via search(), peek(), lines(), etc.
• Only query results (kilobytes) enter the LLM's context—never the full file
• Your RAM is the limit, not the model's context window (with a default 1GB safety cap on action tools)

You can load multiple files or entire repos as separate contexts and query them independently.

A 50MB log file? The LLM sees ~1KB of search results. A 2GB database dump? Same—just the slices you ask for.

By default, Aleph sets a 1GB max file size for action tools to avoid accidental overload, but you can raise it with --max-file-size based on your machine. This cap applies to load_file / read_file; load_context still accepts any size you can supply in-memory.

Installation

pip install "aleph-rlm[mcp]"

After installation, you can automatically configure popular MCP clients:

aleph-rlm install

MCP Server

Run Aleph as an MCP server with:

aleph

Use --enable-actions to allow file and command tools.

Integration

Claude Desktop / Cursor / Windsurf

Add Aleph to your mcpServers configuration:

{
  "mcpServers": {
    "aleph": {
      "command": "aleph",
      "args": ["--enable-actions", "--tool-docs", "concise"]
    }
  }
}

Install the /aleph skill for the RLM workflow prompt:

mkdir -p ~/.claude/commands
cp /path/to/aleph/docs/prompts/aleph.md ~/.claude/commands/aleph.md

Then use it like:

/aleph: Find the root cause of this test failure and propose a fix.

Claude Code

To use Aleph with Claude Code, register the MCP server and install the workflow prompt:

# Register the MCP server
claude mcp add aleph aleph -- --enable-actions --tool-docs concise

# Add the workflow prompt
mkdir -p ~/.claude/commands
cp docs/prompts/aleph.md ~/.claude/commands/aleph.md

Codex CLI

Add to ~/.codex/config.toml:

[mcp_servers.aleph]
command = "aleph"
args = ["--enable-actions", "--tool-docs", "concise"]

How It Works

1. Load: Store a document in external memory via load_context or load_file (with --enable-actions).
2. Explore: Search for patterns using search_context or view slices with peek_context.
3. Compute: Run Python scripts over the content in a secure sandbox via exec_python.
4. Finalize: Generate an answer with linked evidence and citations using finalize.

Recursion: Handling Very Large Inputs

When content is too large even for slice-based exploration, Aleph supports recursive decomposition:

1. Chunk the content into manageable pieces
2. Spawn sub-agents to analyze each chunk
3. Synthesize findings into a final answer

# exec_python
chunks = chunk(100_000)  # split into ~100K char pieces
results = [sub_query("Extract key findings.", context_slice=c) for c in chunks]
final = sub_query("Synthesize into a summary:", context_slice="\n\n".join(results))

sub_query can use an API backend (OpenAI-compatible) or spawn a local CLI (Claude, Codex, Aider) - whichever is available.

Sub-query backends

When ALEPH_SUB_QUERY_BACKEND is auto (default), Aleph chooses the first available backend:

1. API - if API credentials are available
2. claude CLI - if installed
3. codex CLI - if installed
4. aider CLI - if installed

Quick setup:

# OpenAI-compatible API (OpenAI, Groq, Together, local LLMs, etc.)
export ALEPH_SUB_QUERY_API_KEY=sk-...
export ALEPH_SUB_QUERY_MODEL=gpt-5.2-codex

# Optional: custom endpoint
export ALEPH_SUB_QUERY_URL=https://api.your-provider.com/v1

Note: Some MCP clients don't reliably pass env vars from their config to the server process. If sub_query reports "API key not found" despite your client's MCP settings, add the exports to your shell profile (~/.zshrc or ~/.bashrc) and restart your terminal/client.

For a full list of options, see docs/CONFIGURATION.md.

Available Tools

Aleph exposes the full toolset below.

Core exploration

Tool	Description
load_context	Store text or JSON in external memory.
list_contexts	List loaded contexts and metadata.
peek_context	View specific line or character ranges.
search_context	Perform regex searches with surrounding context.
chunk_context	Split content into navigable chunks.
diff_contexts	Diff two contexts (text or JSON).
exec_python	Run Python code over the loaded content.
get_variable	Retrieve a variable from the exec_python sandbox.

Reasoning workflow

Tool	Description
think	Structure reasoning for complex problems.
get_status	Show current session state.
get_evidence	Retrieve collected citations.
evaluate_progress	Self-evaluate progress with convergence tracking.
summarize_so_far	Summarize progress on long tasks.
finalize	Complete with answer and evidence.

Recursion

Tool	Description
sub_query	Spawn a sub-agent on a content slice.

Session management

Tool	Description
save_session	Persist current session to file.
load_session	Load a saved session from file.

Recipes and reporting

Tool	Description
load_recipe	Load an Alephfile recipe for execution.
list_recipes	List loaded recipes and status.
finalize_recipe	Finalize a recipe run and generate a result bundle.
get_metrics	Get token-efficiency metrics for a recipe/session.
export_result	Export a recipe result bundle to a file.
sign_evidence	Sign evidence bundles for verification.

Remote MCP orchestration

Tool	Description
add_remote_server	Register a remote MCP server.
list_remote_servers	List registered remote MCP servers.
list_remote_tools	List tools available on a remote server.
call_remote_tool	Call a tool on a remote MCP server.
close_remote_server	Close a remote MCP server connection.

Action tools

Enabled with the --enable-actions flag. Use --workspace-root and --workspace-mode (fixed, git, any) to control scope.

Tool	Description
load_file	Load a workspace file into a context.
read_file / write_file	File system access (workspace-scoped).
run_command	Shell execution.
run_tests	Execute test commands (supports optional cwd).

Configuration

For full configuration options (limits, budgets, and backend details), see docs/CONFIGURATION.md.

Changelog

Unreleased

• Unlimited context architecture: Clarified that file size is limited by system RAM (with a default 1GB action-tool cap) rather than LLM context windows. Load gigabytes of data and query it with search/peek/lines.
• Added --workspace-mode for action tools (fixed, git, any) to support multi-repo workflows.
• Added optional cwd for run_tests to run tests outside the server’s default working directory.
• Updated MCP setup docs with multi-repo configuration examples.

Development

git clone https://github.com/Hmbown/aleph.git
cd aleph
pip install -e ".[dev,mcp]"
pytest

See DEVELOPMENT.md for architecture details.

References

Aleph implements the Recursive Language Model (RLM) architecture described in:

Recursive Language Models Zhang, A. L., Kraska, T., & Khattab, O. (2025) arXiv:2512.24601

RLMs treat the input context as an external environment variable rather than part of the prompt. This allows models to programmatically decompose inputs, recursively query themselves over chunks, and synthesize results—processing inputs far beyond their native context window.

License

MIT