ToolRecall — Deterministic Tool Cache for LLM Agents

ToolRecall sits between your agent and the OS (or your API provider). On repeat calls it serves cached results from local SQLite instead of re-executing system commands or re-sending requests to the LLM. Caching is deterministic — byte-identical until mtime/TTL expiry — which qualifies every API call for provider prefix-caching discounts (up to 90% at Anthropic/OpenAI).

1 tick instead of 4: A file read normally needs stat → open → read → close. ToolRecall needs only stat (mtime check) — on cache hit the bytes come from memory, bypassing disk entirely.

Zero pip dependencies. Python 3.11+ stdlib only. 76 KB install. One daemon.

pip install toolrecall    # Installs nothing but ToolRecall itself
toolrecall init            # Interactive security setup (default-deny paths)
toolrecall daemon &         # Start cache daemon

Two ways to use (both on by default — no extra command needed):

Requirements: Python 3.11+ (sqlite3, tomllib, json, http.server, urllib from stdlib).

What It Does

ToolRecall intercepts tool calls at the daemon level and returns cached results when inputs haven't changed:

Dynamic commands (git, ls, curl) and state-changing operations always execute live.

Measured effect

In a 13-hour session (Hermes + Gemini 3.1 Pro, 386 messages, 13 project files):

• 89% hit rate (91% file cache): 827 tool calls served from SQLite instead of OS
• 73% fewer file-read tokens at 3× re-read (~204K → ~55K unique)
• ~81% fewer at 10× re-read (~630K → ~55K unique)
• ~20 min less wait time — each cache hit avoids ~1.5s subprocess fork
• Provider prefix-caching becomes reliable: byte-identical payloads qualify for Anthropic/OpenAI's up-to-90% discount on every call

Source: Benchmark

Architecture

  [ Claude Code ]   [ Cursor IDE ]   [ Hermes Agent ]   [ Any LLM Client ]
         \\                |                |               /
          \\               |               |              /
           \\              |               |             /
        +──────────────────────────────────────────────────────────+
        │  Standard stdio MCP   OR   HTTP (OPENAI_BASE_URL proxy) │
        +──────────────────────────────────────────────────────────+
                          │ Unix Domain Socket (Linux/Mac)
                          │ TCP localhost:8568 (Windows)
        +────────────────▼──────────────────────────────────+
        │         ToolRecall Daemon                         │
        │  ┌─────────────────────────────┐                   │
        │  │   In-Memory LRU (Cache)     │                   │
        │  └──────────────┬──────────────┘                   │
        │  ┌──────────────▼──────────────┐                   │
        │  │   SQLite WAL (Persistent)   │                   │
        │  └─────────────────────────────┘                   │
        │  ┌─────────────────────────────┐                   │
        │  │   MCP Server Multiplexer    │                   │
        │  └──────────────┬──────────────┘                   │
        +─────────────────┼──────────────────+
                          │ Lazy-Loaded stdio Subprocesses
        +─────────────────▼──────────────────+
        │ [ Downstream MCP: GitHub / Time ]  │
        +────────────────────────────────────+

The daemon holds everything: the hybrid in-memory LRU + SQLite WAL cache, the MCP Multiplexer (manages subprocesses for external MCP servers), the Forward Proxy (caches full API responses via body hash), and the Security Gate (path allowlist, sensitive file blocklist, cognitive scan).

All agents share one daemon via either:

• MCP Bridge (toolrecall mcp) — the agent connects as an MCP client and uses cached_read, cached_terminal etc.
• Forward proxy (auto-started on :8569) — the agent's API calls go to localhost:8569 instead of api.anthropic.com. The proxy hashes the request body, checks the cache, and on a hit returns the cached response without ever contacting the provider.

See Architecture for the full design.

MCP Multiplexer

When running multiple agents on the same machine (5 Claude Code sessions + 3 Cursor instances), each one normally spawns its own subprocess for every MCP server (GitHub, Postgres, time…). That's 10× the RAM for the same tool.

The daemon's multiplexer shares one subprocess per server across all agents:

• Lazy loading: servers boot on first call, not at daemon start (~0.01s vs ~1.7s per server)
• Idle timeout: inactive subprocesses killed after 15 min (configurable)
• Failure isolation: one server crash doesn't affect others (auto-reconnect, max 3 attempts)
• Secrets: API tokens loaded from ~/.toolrecall/.env, never exposed to the LLM
• Auto-resolution: Server names auto-resolve from the built-in registry — no command/args needed for common servers

All agents connect to one MCP server in their config: toolrecall mcp.

Quick Config Example

# ~/.toolrecall/config.toml
[mcp_multiplex]
servers = ["time", "github", "fetch"]
#  ↑ auto-resolved: time=builtin, github=builtin, fetch=uvx

No [mcp_multiplex.servers_config] section needed for known servers. Custom servers still use the explicit config.

Built-in Servers (zero deps)

External Servers (needs uvx)

See MCP Multiplexer for full configuration details.

When to use: You run 3+ agents simultaneously on the same machine and they share the same MCP tools.
When to skip: Single agent setup — each agent manages its own MCP servers fine.

Security

ToolRecall doesn't prevent prompt injection — it cages the consequences:

• Default-deny path allowlist: Without config, NO paths are readable. toolrecall init prompts for paths interactively.
• Sensitive file blocklist: .env, .ssh/, .pem, .aws/, etc. are blocked even inside allowed paths.
• allow_terminal=false (default): drops all cached_terminal calls into a void.
• os.path.realpath(): catches ../../../etc/shadow traversal before OS is touched.
• Cognitive Pre-Fight: Deterministic regex scan on MCP tool arguments for override instructions, jailbreak tags, exfiltration URLs. Zero LLM, ~0.001ms hot path.
• AST injection check: Parses tool arguments as Python AST — blocks exec(), eval(), __import__() calls.
• Daemon IPC via UDS: No open ports, immune to SSRF.

See Security Architecture for the full trust boundary.

Quick Reference — CLI

toolrecall init            Create default config.toml and .env  [required once]
toolrecall daemon          Start cache daemon (also starts MCP + forward proxy) [required]
toolrecall mcp             Start MCP Bridge (or: mcp list to see registry) [connect any MCP agent]
toolrecall serve           Forward proxy (cache API responses)  [auto-started with daemon; use for custom port]
toolrecall debug           Start debug/demo server (test cached_read/term via curl)
toolrecall status          Cache status and stats               [optional]
toolrecall stats           Detailed cache statistics (JSON)     [optional]
toolrecall invalidate      Clear all caches                     [optional]
toolrecall reset-stats     Reset statistics counters            [optional]
toolrecall nginx           Generate nginx config                [optional]
toolrecall index           Build/update FTS5 knowledge database [optional]
toolrecall index-memory    Index agent memory stores (MEMORY.md, USER.md) [optional]
toolrecall index-dir       Index a directory (e.g. Obsidian)    [optional]
toolrecall config-set      Set a config value                   [optional]
toolrecall shim            Install/uninstall OS-level cache shim (.pth file) [optional]

Agent Integration

Forward proxy (API-level caching)

Cache API responses before they leave your machine. The forward proxy starts automatically with the daemon — no extra command needed. Works with any OpenAI-compatible provider (OpenAI, Anthropic, DeepSeek, OpenRouter, etc.).

toolrecall daemon &                  # also starts forward proxy on :8569
export OPENAI_BASE_URL=http://localhost:8569/v1   # Any OpenAI-compatible SDK
# or override the base URL in your provider config / client init

MCP Bridge (tool-level caching)

ToolRecall registers MCP tools like cached_read, cached_terminal, cached_write, cached_patch. Connect any MCP agent by adding one server:

{
  "mcpServers": {
    "toolrecall": {
      "command": "toolrecall",
      "args": ["mcp"]
    }
  }
}

This single snippet works for Claude Desktop, Claude Code, Cursor, Cline, Windsurf, Continue, and any MCP-compatible agent with zero per-agent variations.

Configuration

TOML (stdlib tomllib) or YAML (optional, requires pyyaml).

# ~/.toolrecall/config.toml (minimal config — toolrecall init creates a full one)
[mcp]
allowed_paths = ["/home/user/projects"]  # Add your project dirs — default-deny!
allow_terminal = false
default_ttl = 60

[mcp_multiplex]
enabled = true
# Server names auto-resolve: time/github/seqthink/fetch = builtin (no deps),
# filesystem/git/memory = external (needs uvx), or override via [mcp_multiplex.servers_config]
servers = ["time", "github", "fetch"]

[nginx]
# nginx is OPTIONAL — only needed if you want HTTPS/SSL in front of the proxy.
# site_name = "toolrecall"
# domain = "example.com"
# ssl = false

TOOLRECALL_* environment variables override TOML.

Uninstall

pip uninstall toolrecall
python3 scripts/uninstall.py --force

Removes: daemon, systemd service, config, cache DB, logs.

Platform Support

Documentation

• Architecture — daemon design, layers, IPC
• Architecture Diagram — system and sequence diagrams, token costs, Context Tracker
• CLI Reference — all subcommands explained
• Configuration Reference — config.toml, config.py, all env vars
• Context Tracker — checkpoint-based dirty-file tracking, O(n²) breakdown
• How It Works — quick technical overview
• MCP Multiplexer — single-daemon MCP management, server registry
• Testing Guide — test philosophy, organization, per-file coverage
• Benchmark — measured performance, token savings
• Knowledge DB — FTS5 indexing guide
• Docker Deployment — containerized stack
• Security Architecture — WAF details, trust boundary
• Troubleshooting — common fixes
• Appendix — comparison tables, OSI model, ROI, vision, audit
• Hermes Transparent Cache — auto-patching for Hermes Agent

[^notall]: Not all agents tested yet — please report bugs.