Tools

The tools we use, and why.

New AI engineering tools launch every month. We test them on real codebases so your team doesn't have to. This page lists what we use today, what each tool is for and where it runs.

Updated 26 October 2026

No vendor pays us.

We take no referral fees, equity or advisory roles from any tool on this page.

A tool is here because it works on real teams. When something better comes along, we switch, and this page changes with it. If we ever take money from a vendor, we'll say so here first.

One thing to declare. aikaara built Parley, the agent platform behind TaxBuddy's AI agents. It is our own product, so it isn't on this list.

What we use today

Coding agents and editors
  • Claude Code

    AnthropicIn our stack

    What it does: Anthropic's coding agent. It works across a whole codebase from the terminal.

    Why we use it: It handles large changes across many files, and its work lands as pull requests a person reviews.

  • Codex

    OpenAIIn our stack

    What it does: OpenAI's coding agent, in the terminal, the editor and the cloud.

    Why we use it: It can run several tasks in parallel, and each one comes back as a separate change to review.

  • OpenCode

    opencode.aiIn our stack

    What it does: An open-source coding agent for the terminal that works with many model providers.

    Why we use it: Teams can use the models they already pay for, including ones in their own cloud account.

  • Cursor

    cursor.comIn our stack

    What it does: A code editor with AI built in.

    Why we use it: Engineers who want AI inside their editor get it without changing how they work.

Orchestration and review
  • Alan

    tryalan.aiIn use with clients

    What it does: Coordinates coding agents, tests, reviews and releases in one place.

    Why we use it: A person approves every change before it ships. That is the core of our procedure, and Alan enforces it.

    Where it runs: TaxBuddy, International Wheel & Tire and Centrum.

  • GitHub

    github.comIn our stack

    What it does: Where the code, pull requests and Actions live, with Copilot for engineers who use it.

    Why we use it: Our procedure runs through pull requests. Every agent change lands as one, and a person merges it.

AI code review

Review is where AI-written code gets stuck. These tools take the first pass, so your reviewers spend their time on what matters.

  • CodeRabbit

    coderabbit.ai

    Reviews every pull request and comments line by line.

  • Greptile

    greptile.com

    Reviews pull requests with context from the whole codebase.

  • Graphite

    graphite.dev

    Stacked pull requests and AI review, so changes stay small and quick to approve.

Tracing and evals

Agents in production need checking like any other system. These show what an agent did, and test it before a change goes out.

  • Langfuse

    langfuse.com

    Open-source tracing, prompt management and evals.

  • Braintrust

    braintrust.dev

    Evals and monitoring for AI products.

  • LangSmith

    langchain.com

    Tracing and evals from the LangChain team.

AI spend tracking and routing

This is how we keep AI spend under control. Every model call goes through a gateway that records its cost, applies team budgets and sends routine work to cheaper models.

  • LiteLLM

    litellm.ai

    An open-source gateway to many model providers, with budgets per team and key.

  • OpenRouter

    openrouter.ai

    One API for many models, with usage and cost per key.

  • Helicone

    helicone.ai

    Logs every model call with its cost and latency.

  • Portkey

    portkey.ai

    A gateway with budgets, rate limits and routing across providers.

Frontier models

We use all the frontier models, and pick the one that does each job best. Hard problems, large changes and reviews go to these.

  • Claude

    Anthropic

    Strong on large codebases, long tasks and careful reasoning.

  • GPT

    OpenAI

    Broad, capable models across coding, reasoning and tool use.

  • Gemini

    Google

    Long context and multimodal work, inside Google Cloud when you need it.

Fast, low-cost models

Routine agent work runs on fast, low-cost models. These are the ones we run most today.

  • DeepSeek V4.1 Flash

    DeepSeek

    Fast, low-cost model for everyday coding and agent tasks.

  • DeepSeek V4 Flash DSpark

    DeepSeek

    V4 Flash with built-in speculative decoding, so answers come back faster. Open weights, so it can run on your own GPUs.

  • GLM 5.2

    Z.ai

    General model for coding and agent work.

  • GLM 5.3 Flash

    Z.ai

    Faster, lighter GLM for high-volume tasks.

New models come out every few weeks. We test each one on real work as it lands, and move work to it when it does the job better or for less. This list changes with them.

Compute, CI and sandboxes
  • Modal

    modal.comIn our stack

    What it does: GPU and compute on demand.

    Why we use it: Agent and model workloads start in seconds, and you pay only while they run.

  • Blacksmith

    blacksmith.shIn our stack

    What it does: Faster runners for GitHub Actions.

    Why we use it: Agents open more changes than people do. Slow CI becomes the bottleneck, and faster runners keep tests from holding up review.

  • E2B

    e2b.devIn our stack

    What it does: Secure cloud sandboxes for running code.

    Why we use it: Agents run and test code in an isolated sandbox, away from your systems and data.

Cloud AI platforms
  • Azure AI Foundry

    azure.microsoft.comYour choice of cloud

    What it does: Microsoft's platform for models and agents inside Azure.

    Why we use it: Many Microsoft shops already have it approved by security and procurement.

  • AWS Bedrock

    aws.amazon.comYour choice of cloud

    What it does: Amazon's managed access to foundation models inside AWS.

    Why we use it: Teams on AWS can keep models and data in the account they already run.

  • Google Vertex AI

    cloud.google.comYour choice of cloud

    What it does: Google's platform for models and agents inside Google Cloud.

    Why we use it: Teams on Google Cloud can use it without a new vendor contract.

We work inside whichever one your company already uses. Your models and data can stay in your own cloud account.

How a tool gets on this list

  1. We see the demo.I sit through a lot of them. Most tools stop here.
  2. We run it on our own code first.A tool has to hold up on real work before it goes near a client.
  3. It goes to a client team.It has to fit the procedure. Agents propose and a person still approves.
  4. We keep checking it.If it stops earning its place, we take it off and say why.

Building a tool for engineering teams?

We work as engineering and design partners to the companies that make these tools. We run your tool on real teams and tell you plainly what breaks. You get to value faster, and so do our clients.

What you get

  • Your tool on real codebases, with real deadlines
  • Fast, specific feedback from the engineers using it
  • A partner who helps shape the product

What we won't do

  • Take a fee, equity or an advisory role
  • List a tool that hasn't held up in real work
  • Recommend you to a client it doesn't suit

Want these tools working on your code?