aleph-rlm 0.8.7

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

Aleph

Aleph is an MCP server plus companion skill workflow (/aleph in Claude Code, $aleph in Codex CLI) for recursive LLM work. It stores working data in a Python process and exposes tools so the model can retrieve slices, run code, and iterate without repeatedly injecting full files into prompt context.

Core capabilities:

• Load large files and codebases into process memory
• Search and inspect targeted ranges (search_context, peek_context)
• Run computation over context with exec_python
• Orchestrate recursive sub-queries and recipe pipelines
• Save and restore sessions for long investigations

Design is based on the Recursive Language Model (RLM) architecture.

+-----------------+    tool calls     +--------------------------+
|   LLM client    | ---------------> |  Aleph (Python process)  |
| (context budget)| <--------------- |  search / peek / exec    |
+-----------------+   small results  +--------------------------+

Quick Start

1. Install:

pip install "aleph-rlm[mcp]"

2. Auto-configure your MCP client:

aleph-rlm install

aleph-rlm install / aleph-rlm configure now treats Codex as the default sub-query path. When the codex CLI is installed, generated configs pin the Codex MCP defaults (backend=codex, mode=mcp, model=gpt-5.4, reasoning=low, share_session=true) unless you override them. Other CLI backends remain available as explicit experimental overrides.

Recommended setup:

• Best default: install Codex CLI, then run aleph-rlm install. This is the simplest and strongest path even if your top-level client is Claude Code. Aleph can keep Claude as the outer client and still use Codex for nested sub-queries.
• All-Claude alternative: if you want Claude to handle nested sub-queries too, explicitly set ALEPH_SUB_QUERY_BACKEND=claude and ALEPH_SUB_QUERY_SHARE_SESSION=true, or pick claude in aleph-rlm configure.

3. Verify Aleph is reachable in your assistant:

get_status()
# or
list_contexts()

4. Run the skill flow on a real file:

/aleph path/to/large_file.log
# or in Codex CLI
$aleph path/to/large_file.log

Expected behavior: Aleph loads the file into process memory, then begins analysis with tool calls (search_context, peek_context, exec_python) without requesting pasted raw content.

Two-Minute Smoke Test

Use a file inside your current Aleph workspace and run:

get_status()
load_file(path="/absolute/path/to/file", context_id="smoke")
search_context(pattern="TODO|ERROR|WARN", context_id="smoke")
exec_python(code="result = {'lines': line_count(), 'headings': len(search(r'^## '))}", context_id="smoke")
get_variable(name="result", context_id="smoke")
save_session(context_id="smoke", path=".aleph/smoke_session.json")

Healthy behavior:

1. get_status() reports the expected workspace_root
2. load_file(...) creates the context without pasting raw file content
3. search_context, exec_python, and get_variable return compact derived results
4. save_session(...) succeeds when the path stays under the workspace root

If you try to save outside the workspace root, Aleph rejects the path instead of silently writing elsewhere.

Common Workloads

Scenario	What Aleph Does
Large log analysis	Load large logs, trace patterns, correlate events
Codebase navigation	Search symbols, inspect routes, trace behavior
Data exploration	Analyze JSON/CSV exports with Python helpers
Mixed document ingestion	Load PDFs, Word docs, HTML, and compressed logs
Semantic retrieval	Use semantic search, then zoom with line/char peeks
Long investigations	Save sessions and resume from memory packs

Commands

Installing aleph-rlm gives you three commands:

Command	Purpose
aleph	MCP server (also supports run / shell)
aleph-rlm	Installer/config helper (also supports run / shell)
alef	Legacy standalone CLI (deprecated)

How to think about it:

• Run aleph-rlm install once to configure clients.
• MCP clients should run aleph as the server command.
• Use aleph run (or aleph-rlm run) for terminal-only mode.

MCP Mode

Automatic Setup

aleph-rlm install

To customize workspace scope, backend, docs mode, or Docker settings:

aleph-rlm configure

Manual Setup (Any MCP Client)

Use this as a practical default:

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

Verify MCP Wiring

In your assistant session:

get_status()

If your client namespaces tools, use mcp__aleph__get_status.

Config File Locations

Client	macOS/Linux	Windows
Claude Code	~/.claude/settings.json	%USERPROFILE%\.claude\settings.json
Claude Desktop	~/Library/Application Support/Claude/claude_desktop_config.json	%APPDATA%\Claude\claude_desktop_config.json
Cursor	~/.cursor/mcp.json	%USERPROFILE%\.cursor\mcp.json
VS Code	~/.vscode/mcp.json	%USERPROFILE%\.vscode\mcp.json
Codex CLI	~/.codex/config.toml	%USERPROFILE%\.codex\config.toml

More per-client setup details are in MCP_SETUP.md.

The /aleph and $aleph Skill

For skill-based usage, configure both:

1. MCP server configured in the client
2. Skill prompt installed (docs/prompts/aleph.md)

Invocation

Client	Skill command	Typical usage
Claude Code	/aleph	/aleph path/to/file
Codex CLI	$aleph	$aleph path/to/file

Skill Install Paths

Download docs/prompts/aleph.md and place it at:

• Claude Code: ~/.claude/commands/aleph.md
• Codex CLI: ~/.codex/skills/aleph/SKILL.md

Windows equivalents:

• %USERPROFILE%\.claude\commands\aleph.md
• %USERPROFILE%\.codex\skills\aleph\SKILL.md

Quick Behavior Check

Use this exact prompt:

$aleph path/to/large_file.log
Then call list_contexts() and show the loaded context_id before analysis.

Healthy behavior:

1. Tool call to load_file(path=...)
2. Context appears in list_contexts()
3. Follow-up search/peek/exec on that context

Backend Selection Rules

Aleph resolves the active sub-query backend in this order:

1. Programmatic config via configure(sub_query_backend=...) or SubQueryConfig(backend=...)
2. ALEPH_SUB_QUERY_BACKEND when it is set to a concrete backend
3. Auto-detection: codex -> api

That means an explicit runtime switch wins over ambient shell state, while auto keeps the first-class Codex path and otherwise falls back to API.

Shared-Session Architecture

When ALEPH_SUB_QUERY_SHARE_SESSION=true, Aleph starts a local streamable HTTP server and points the sub-query CLI back at that live Aleph session. That is what lets nested sub-agents use search_context, peek_context, exec_python, and the rest of Aleph's MCP tools instead of relying on a pasted prompt slice.

How Aleph injects that live MCP server depends on the backend:

Backend	How shared-session MCP is injected
codex	Native Codex MCP config overrides via codex mcp-server
claude	Temp JSON file via --mcp-config and --strict-mcp-config
gemini	Temp JSON file via GEMINI_CLI_SYSTEM_SETTINGS_PATH

This is why Codex is the recommended default: it has the cleanest nested MCP path, the least ambient config leakage, and the best retry/thread behavior.

Core Workflow Patterns

1) Load File -> Work Immediately

load_file(path="/absolute/path/to/large_file.log", context_id="doc")
search_context(pattern="ERROR|WARN", context_id="doc")
peek_context(start=1, end=60, unit="lines", context_id="doc")
exec_python(code="print(line_count())", context_id="doc")
finalize(answer="Summary...", context_id="doc")

Note: with MCP action tools, absolute paths are safest for load_file.

2) Analyze Raw Text

load_context(content=data_text, context_id="doc")
search_context(pattern="keyword", context_id="doc")
finalize(answer="Found X at line Y", context_id="doc")

3) Recipe Pipelines

Recommended sequence:

validate_recipe -> estimate_recipe -> run_recipe

Example:

run_recipe(recipe={
  "version": "aleph.recipe.v1",
  "context_id": "doc",
  "budget": {"max_steps": 6, "max_sub_queries": 5},
  "steps": [
    {"op": "search", "pattern": "ERROR|WARN", "max_results": 10},
    {"op": "map_sub_query", "prompt": "Root cause?", "context_field": "context"},
    {"op": "aggregate", "prompt": "Top causes with evidence"},
    {"op": "finalize"}
  ]
})

4) Sub-Query Batching (Important)

Prefer fewer large sub-query calls over many tiny calls.

• Bad: 1000 calls of 1K chars
• Good: 5-10 calls of about 100K to 200K chars

exec_python(code="""
chunks = chunk(100000)
summaries = sub_query_batch("Summarize this chunk:", chunks)
print(summaries)
""", context_id="doc")

5) Save and Resume

save_session(context_id="doc", path=".aleph/session_doc.json")
load_session(path=".aleph/session_doc.json")

Paths for save_session() / load_session() must stay inside Aleph's workspace root. .aleph/ under the repo is a safe default.

CLI Mode (Standalone)

Use this when you want Aleph without MCP integration.

# Basic
aleph run "What is 2+2?" --provider cli --model codex

# With file context
aleph run "Summarize this log" --provider cli --model codex --context-file app.log

# JSON output with trajectory
aleph run "Analyze" --provider cli --model codex --context-file data.json --json --include-trajectory

Common Flags

Flag	Description
--provider cli	Use local CLI tools instead of API provider
`--model codex	claude
--context-file <path>	Load context from file
--context-stdin	Read context from stdin
--json	Emit JSON output
--include-trajectory	Include full reasoning trace
--max-iterations N	Limit loop steps

Common Environment Variables

Variable	Description
ALEPH_SUB_QUERY_BACKEND	auto, codex, gemini, kimi, claude, or api
ALEPH_SUB_QUERY_TIMEOUT	Sub-query timeout in seconds
ALEPH_SUB_QUERY_SHARE_SESSION	Share MCP session with CLI sub-agents
ALEPH_SUB_QUERY_CODEX_MODE	Codex backend mode: defaults to mcp
ALEPH_SUB_QUERY_CODEX_MODEL	Codex MCP model override, default gpt-5.4
ALEPH_SUB_QUERY_CODEX_REASONING_EFFORT	Codex MCP reasoning effort, default low
ALEPH_SUB_QUERY_GEMINI_SANDBOX	Opt back into Gemini CLI sandboxing (false by default for Aleph sub-queries)
ALEPH_CLI_TIMEOUT	Timeout for CLI calls

When Aleph resolves to the codex backend, it now defaults to the internal codex mcp-server path with a clean MCP graph (-c mcp_servers={}).

gemini, claude, and kimi remain available only when you explicitly pin them via ALEPH_SUB_QUERY_BACKEND or runtime configure(...).

In live dogfooding, claude and gemini both worked as shared-session Aleph MCP backends after explicit selection, but Codex was more reliable for exact-output and retry-sensitive tasks.

Aleph launches Gemini sub-queries with --extensions "" so the nested run does not inherit unrelated user extensions from ~/.gemini.

If you want the simplest Codex-backed shared-session setup:

aleph-rlm install

If you want an all-Claude setup instead:

export ALEPH_SUB_QUERY_BACKEND=claude
export ALEPH_SUB_QUERY_SHARE_SESSION=true

If Codex is installed, aleph-rlm install will still pin Codex by default until you override it.

Use these env vars if you want to pin the Codex defaults explicitly:

export ALEPH_SUB_QUERY_BACKEND=codex
export ALEPH_SUB_QUERY_CODEX_MODE=mcp
export ALEPH_SUB_QUERY_CODEX_MODEL=gpt-5.4
export ALEPH_SUB_QUERY_CODEX_REASONING_EFFORT=low
export ALEPH_SUB_QUERY_SHARE_SESSION=true

Quick smoke for the nested Codex MCP path:

load_context(content="Fact: nested codex mcp smoke works", context_id="smoke")
exec_python(code="result = sub_query(\"Use Aleph MCP tools to inspect context_id='smoke'. Return exactly the fact text and nothing else. Do not rely on embedded context.\")", context_id="smoke")
get_variable(name="result", context_id="smoke")

If get_variable(...) returns nested codex mcp smoke works exactly, Aleph is sharing the live MCP session with the nested Codex agent and the nested agent is reading through Aleph tools instead of prompt-stuffed context.

Tool Overview

Core Tools (Always Available)

Category	Tools
Context	load_context, list_contexts, diff_contexts
Search	search_context, semantic_search, peek_context, chunk_context
Compute	exec_python, get_variable
Reasoning	think, evaluate_progress, summarize_so_far, get_evidence, finalize
Runtime Config	configure
Recipes	validate_recipe, estimate_recipe, run_recipe, compile_recipe, run_recipe_code

Action Tools (--enable-actions)

Category	Tools
Filesystem	load_file, read_file, write_file
Shell	run_command, run_tests, rg_search
Persistence	save_session, load_session
Remote MCP	add_remote_server, list_remote_servers, list_remote_tools, call_remote_tool, close_remote_server

exec_python includes 100+ helpers (search, chunk, lines, extract_*, sub_query, sub_query_batch, sub_query_map, sub_aleph, Recipe DSL helpers, and more). Recursion helpers are available inside exec_python, not as top-level MCP tools.

Swarm Mode (Optional)

Aleph can act as shared memory for multiple agents.

Agent A/B/C <-> Aleph contexts in shared RAM

Simple pattern:

1. Shared KB context: swarm-<name>-kb
2. Task contexts: task-<id>-spec, task-<id>-findings
3. Agent-private contexts: <agent>-workspace

Example write/read:

exec_python(code="ctx_append('Auth uses JWT with RS256')", context_id="task-42-findings")
search_context(pattern="JWT", context_id="task-42-findings")

Context Isolation and Safety

Aleph enforces strict boundaries to prevent raw context from leaking into the LLM's context window:

• System prompt isolation. The default system prompt does not include a raw context preview. The placeholder is replaced with [OMITTED FOR CONTEXT ISOLATION].
• get_variable("ctx") is policy-aware. In isolated policy, retrieving ctx via the MCP boundary is blocked with guidance. In trusted policy, it is allowed but still subject to response caps/truncation. Prefer processing data inside exec_python and retrieving compact derived results with get_variable.
• Execution output truncation. exec_python stdout, stderr, and return values are all truncated to max_output_chars (default 50,000). The MCP tool response is further capped at max_tool_response_chars (default 10,000). Both limits are configurable.
• Tool response caps. Every MCP tool response (peek, search, semantic search, get_variable, etc.) is bounded by the same response-size cap.

Deployment Profiles

Set ALEPH_CONTEXT_POLICY to choose a profile:

Profile	Behavior
trusted (default)	Low friction. Auto memory-pack, session save/load without confirmation.
isolated	Explicit consent. Requires confirm=true for session export/import, disables auto memory-pack. Blocked tools return actionable alternatives.

Switch at runtime with configure(context_policy="isolated"). See CONFIGURATION.md for details.

Safe Usage Pattern

# Compute server-side — data stays in Aleph RAM
exec_python(code="""
errors = [l for l in ctx.splitlines() if 'error' in l.lower()]
result = f'Found {len(errors)} errors. First 3: {errors[:3]}'
""", context_id="doc")

# Retrieve only the small derived result
get_variable(name="result", context_id="doc")

Avoid returning full-context payloads unless necessary. In isolated policy, get_variable("ctx") is blocked; in trusted policy large raw responses are still truncated by output caps.

Configuration Quick Reference

Workspace and Safety

Flag/Variable	Purpose
--workspace-root <path>	Root for relative action paths
`--workspace-mode <fixed	git
--require-confirmation	Require confirm=true for actions
ALEPH_WORKSPACE_ROOT	Override workspace root
ALEPH_CONTEXT_POLICY	trusted (default) or isolated
ALEPH_OUTPUT_FEEDBACK	full (default) or metadata
ALEPH_MAX_RECIPE_CONCURRENCY	Max parallel map_sub_query tasks (default 10)

Limits

Flag	Default	Purpose
--max-file-size	1 GB	Max file read size
--max-write-bytes	100 MB	Max file write size
--timeout	180 s	Sandbox/command timeout
--max-output	50,000 chars	Max command output
ALEPH_MAX_TOOL_RESPONSE_CHARS	10,000 chars	MCP tool response cap

Recursion Budgets

Variable	Default	Purpose
ALEPH_MAX_DEPTH	2	Max sub_aleph nesting depth
ALEPH_MAX_ITERATIONS	100	Total RLM steps
ALEPH_MAX_WALL_TIME	300 s	Wall-time cap
ALEPH_MAX_SUB_QUERIES	100	Max sub_query calls
ALEPH_MAX_TOKENS	unset	Optional per-call output cap

Full configuration details: docs/CONFIGURATION.md

Troubleshooting

• Tool not found: ensure Aleph MCP server is running.
• Context not found: verify context_id and check list_contexts().
• No search hits: broaden regex or use semantic_search.
• rg_search is slow: install ripgrep (rg).
• Running out of context: use summarize_so_far().
• Session load errors: check file path and memory pack schema.

Documentation

Document	Purpose
MCP_SETUP.md	Client-by-client MCP configuration
docs/CONFIGURATION.md	Full flags and environment variables
docs/langgraph-rlm-default.md	LangGraph integration with RLM-default tool usage
examples/langgraph_rlm_repo_improver.py	Repo-improvement runner with optional LangSmith tracing
docs/prompts/aleph.md	Skill workflow and tool reference
CHANGELOG.md	Release history
DEVELOPMENT.md	Contributor guide

Development

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

References

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

License

MIT