toolahead

Stop waiting for tools

Agents normally work serially:

reason → call tool → wait → inspect → reason → call tool → wait

ToolAhead learns which calls usually follow each other. It starts the likely next call while the model is still working:

Agent       inspect result ───── reason ───── request next tool ── result
ToolAhead                  └──── run predicted tool ──────────────┘

The agent still calls ordinary MCP tools. If no matching result is ready, the tool runs normally. If ToolAhead prepared the exact call against the exact same files, the result returns immediately from memory.

Related MCP server: Token-Efficient MCP Server

Real agents. Real tools. Work already done.

Codex: the same task with and without ToolAhead

This is a 1× recording of the third matched test pair—not a scripted demo. Both runs used the real Codex API and separate copies of the same project. ToolAhead prepared all eight predicted calls, reused both 7.5-second test results, and finished the same task 19.93 seconds earlier.

20 real-agent runs with multiple edits

For each agent, five runs with ToolAhead were paired with five runs without it. Every run performed the same 12-call workflow: list, search, four reads, failing tests, one intentionally rejected edit, three successful file changes, and final tests. ToolAhead started each accelerated run with patterns learned from the same tool sequence.

All correct runs are included, even when the accelerated run took longer overall. With only five pairs per agent, the statistical ranges are wide: the estimated median improvement is 2.73–26.17 seconds for Codex and −8.25–37.98 seconds for Claude. The controlled test command takes 7.5 seconds, so the seconds of tool work hidden in the background are more useful than a percentage that changes with every workload.

Claude shows the difference between faster tools and a faster complete task: ToolAhead removed 9.42–10.89 seconds of tool waiting in every accelerated run, but Claude's API and model response times varied enough to dominate the five overall comparisons. The paired results and statistical ranges remain in the benchmark data.

The complete protocol, paired measurements, correctness checks, and statistical range are summarized in BENCHMARKS.md.

Full recorded runs

These 1× recordings show ToolAhead handling the complete workflow: list, search, read, edit, write, test, result validation, and reuse.

Codex CLI

Claude Code

Install

Once published on PyPI:

uvx toolahead --help
# or
python3 -m pip install toolahead

From a local checkout today:

git clone https://github.com/michael-ra/toolahead.git
cd toolahead
uvx --from . toolahead --help

Requirements: Python 3.11+, macOS or Linux, and an authenticated Codex CLI or Claude Code installation. watchdog is optional.

Quickstart

Run these commands inside the project you want to accelerate:

# Connect both agents to ToolAhead and install the required hooks.
uvx toolahead init --agent both --strict --project .

# Allow this exact test command to run ahead and be reused.
uvx toolahead allow "python3 -m pytest" --project .

# Start ToolAhead in the background for this workspace.
uvx toolahead serve --workspace .

Then start your agent in a second terminal.

Codex CLI:

codex

Claude Code:

ANTHROPIC_BASE_URL=http://127.0.0.1:4242 claude

See live timing and cache statistics at any time:

uvx toolahead status

Rerun toolahead init after upgrading ToolAhead. It refreshes ToolAhead's project files without changing unrelated Codex, Claude, or MCP settings.

One clear set of tools

ToolAhead gives the agent one consistent set of MCP tools. This lets it return prepared results directly instead of waiting for a native tool to run and then trying to replace its result afterward.

The agent never sees cache wrappers or duplicate JSON. ToolAhead keeps cache timing in hidden MCP _meta; prepared and normal calls return the same text, errors, and exit codes.

Why --strict matters

Showing two equivalent Read tools forces the model to choose between duplicate options, wastes prompt space, and makes selection less reliable. Strict mode keeps one set:

• Claude Code's project settings hide native Read, Grep, Glob, Edit, and Write; the six ToolAhead MCP equivalents take their place.

• Codex sees the same six tools and instructions to use them. Strict mode redirects native apply_patch to edit_file so edit→test learning stays intact. Codex's general shell remains available when needed; explicitly allowed Bash tests can still reuse prepared results.

• Tool names and field conventions stay close to the native coding-agent tools. Descriptions are intentionally short to reduce the tokens sent to the model.

Omit --strict if you want to keep all native file tools visible while trying ToolAhead.

Predictions can be wrong. Returned results cannot.

ToolAhead is free to guess what comes next, but it returns prepared work only when the requested call and current files are exact matches.

flowchart LR
    A[Previous tool or turn start] --> B[Predict next exact call]
    B --> C[Read-only worker or disposable checkout]
    A --> D[Agent keeps reasoning]
    C --> E{Exact call + fresh SHA-256 input match?}
    D --> E
    E -->|match| F[Return prepared result from RAM]
    E -->|no match| G[Execute the MCP call normally]

• List, Search, and Read results are tied to the exact request and the relevant file contents.

• Command results are tied to the exact command and a fresh hash of the whole workspace.

• Commands run ahead only in a disposable workspace copy.

• A prepared result is returned only when the real workspace still matches the copy used to create it.

• Wrong predictions, background-process failures, expired results, and timeouts automatically fall back to a normal tool execution.

• Cache entries store stdout, stderr, and exit code—not a model-generated summary.

ToolAhead learns tool sequences locally. The reliable signal is the previous tool finishing; visible commentary can offer an earlier hint when an agent provides it. Private chain-of-thought is never required.

Latest file change wins

ToolAhead does not need to guess which edit will be the last one. Every successful Edit or Write increases a simple workspace version number:

edit version 1 ── start predicted tests
edit version 2 ── stop version 1 ── restart tests on version 2
edit version 3 ── stop version 2 ── keep only the version 3 result

• A running command for an older file version receives SIGTERM as a process group, then SIGKILL if it does not stop promptly.

• The pending command is restarted for the newest file version even when another edit arrives before the test request.

• Writes arriving within 50 ms are grouped before work starts. Configure the window with PREFETCH_MUTATION_DEBOUNCE_MS; set it to 0 to disable grouping.

• Outdated results are never inserted into the current cache. Fresh SHA-256 validation remains the final replay condition.

• Failed file changes do not increase the workspace version.

In plain terms: after every successful file change, ToolAhead starts the likely next safe call. Nearby changes are grouped, and a newer change always replaces work started for an older file state.

Which commands can be reused

ToolAhead may return a prepared command result instead of running the command again only when that exact command is listed in .prefetch-replay.json:

{
  "commands": [
    "python3 -m pytest",
    "npm test"
  ]
}

Use the CLI instead of editing the file by hand:

toolahead allow "python3 -m pytest" --project .

The allowed-command list updates without restarting ToolAhead. It rejects shell chains, pipes, redirects, substitutions, installers, and arbitrary commands; recognized test/lint families include unittest, pytest, npm/yarn tests, Go, Cargo, Make, Jest, Vitest, Ruff, ESLint, TypeScript, and mypy.

Prioritize known failures without weakening the result

Use the test runner's explicit full-suite mode when available. For pytest, pytest --ff runs the last failures first and then the rest of the suite; ToolAhead can learn and reuse that exact command normally. Focused modes such as pytest --lf or Jest --onlyFailures are useful quick checks, but ToolAhead never substitutes their partial result for a requested full-suite result.

Latency metrics

toolahead status separates the parts that can otherwise be confused:

This is why removing 5 seconds of tool waiting does not guarantee the complete task finishes exactly 5 seconds sooner: model and API response times vary independently.

Security model

• Tool paths are contained inside the configured workspace; symlink escapes are rejected.

• Every command run ahead uses a fresh disposable copy, never the live checkout.

• The local daemon binds to 127.0.0.1 and adds no remote telemetry.

• Before returning a prepared result, ToolAhead hashes the current files again. Filesystem watchers only help it skip unnecessary hashing.

• Tests that depend on external services, databases, clocks, random values, or environment state cannot be validated from source files alone.

Limitations

• Edit and Write are intentionally not run ahead. After either finishes, ToolAhead starts the next predicted safe tool. Rapid changes are grouped, and commands running against an older file state are stopped.

• Prepared command results are limited to explicitly approved tests, builds, and linters whose output should be repeatable.

• Commands currently verify the entire workspace, which can be conservative on very large monorepos. Checking only relevant dependencies is planned.

• API and model response times can outweigh the saved tool time. Compare multiple runs with and without ToolAhead instead of relying on one attempt.

• Hosted tools such as provider-side web search cannot be run ahead by this local integration.

• Windows has not yet been validated.

Development

Build and verify the PyPI artifacts:

uv build
python3 .github/scripts/normalize_sdist.py dist/*.tar.gz
python3 .github/scripts/check_distribution.py dist/*.whl dist/*.tar.gz
uvx --from twine twine check dist/toolahead-0.2.0a2*
uvx --from dist/toolahead-0.2.0a2-py3-none-any.whl toolahead --help

Project map

• src/toolahead/ — installable CLI, MCP server, prediction engine, hooks, sandbox execution, replay, and telemetry

• docs/assets/ — the logo and README recordings

• .github/workflows/ — package validation and trusted PyPI publishing

• .github/scripts/ — release-archive privacy and metadata checks

Research foundations

ToolAhead is an independent implementation informed by research on speculative tool execution. It is not an official implementation or reproduction of any single paper. The closest foundations are:

• SPORK: Self-Speculative Forking to Accelerate Agentic LLM Inference and Parallelizing Tool Execution and LLM Generation (PASTE) for overlapping predicted tool execution with ongoing agent reasoning.

• Speculate with Memory for learning recurring action transitions from previous agent trajectories.

• SpecBox for speculative sandbox prewarming and isolated execution.

• AOSpec for lossless action/state verification across speculative execution.

ToolAhead combines these directions with local Codex and Claude Code hooks, exact call-and-workspace matching, MCP result replay, mutation generations, and a standalone Python package. All benchmark numbers above are ToolAhead's own measurements, not results reported by those papers.

License

Apache License 2.0. See LICENSE.

Contributions are welcome; see CONTRIBUTING.md. Security reports should follow SECURITY.md.