AskCodi
Orchestrated AI for developers: a Mac app and an API
Start new thread

AskCodi - Orchestrate agents at scale while reducing cost

by
•
Tell Codi what you want to build. It writes the plan, runs AI agents in parallel across all your projects, and picks the cheapest model that can do each task, so you ship more and spend less. When you add a project it writes the charter and task list itself, and only pulls you in when a decision needs you, through one inbox.

Add a comment

Replies

Best
Maker
📌
Hi, I'm Shreyans, one of the makers. I built AskCodi because I got tired of babysitting AI. Every tool I tried turned building software into a second job. Pick the model, paste the context, wire up the agents, then sit there watching a run and hoping it didn't go sideways. I was spending more time operating the tool than shipping with it. So we made the opposite of that. You hire Codi, an AI CTO that lives on your Mac, and you talk to it the way you'd talk to a real one. You don't manage a fleet of agents. You manage one person, and that person runs the team. You've probably seen tools that give you a control center for your agents: terminals, panes, worktrees, all of it yours to drive. This is the opposite. I didn't want a bigger cockpit. I wanted to hand the work to someone and step away. A few things I'm proud of, and would love you to poke holes in. Before Codi writes a single line, it sits down and learns the project. It asks what you're building, who it's for, and what it should never touch without checking, then drafts a plan from your answers. So it builds the thing you meant, not the literal words you typed. You never choose a model again. Codi decides task by task. Cheap models for the boring parts, the expensive ones only where they earn it. In our own use, a week of work came out around a dollar, instead of the fifteen-ish it costs to run a frontier model on everything. And it doesn't need you sitting over it. Codi already knows what's next from talking to you, so it keeps working in your off-hours when tokens are cheaper. You close the laptop, it keeps going, and the calls that are actually yours are waiting in an inbox the next morning. That's really the whole product: a chat and an inbox. It also isn't one thing at a time. Codi runs sessions in parallel across every project you have, so you can kick off work on three repos before your coffee is ready and let them all move together. It's the closest I've felt to running a studio on my own. Honest caveat, since I'd rather you hear it from me: it's early. The memory gets sharper the more you use it, and there are rough edges we're still filing down. If you try it, I most want to hear where it annoyed you. One aside: AskCodi also has an OpenAI-compatible API, if you'd rather build on the orchestration than talk to it. The app is what I'm here to show you today, but the same brains are a call away. It's on Mac(Windows coming soon), free to start on your own Claude subscription. Ask me anything, I'll be here all day.

 Congtrats on an awesome launch! Excited to try the new version of AskCodi.

Both routers and sandboxed openclaws are all the rage atm. Seems like you've put them together in a single, inbox-style interface, which seems really cool. But where does it go from here? Can this really become like an AI CTO, that designs your infra, picks the tools, innovates on project concepts, etc?

Maker

 We're already moving AskCodi toward becoming a true AI CTO. Instead of managing prompts or implementation details, you communicate with Codi the way a CEO communicates with a CTO: define the vision, explain customer needs, and describe the outcome you want. From there, state-of-the-art LLMs determine the best path to bring that vision to life.

These days, most of what we do is review and approve with a "yes, yes, yes." It's YOLO and tokenmaxxing combined with the efficiency and guardrails of human oversight.

 The babysitting line is so real. Love that you went the opposite direction from a bigger cockpit. What happens when two projects need a decision from you at the same time, does the inbox stack them up?

Maker

 Yes currently inbox acts exactly as an email inbox. Messages keep stacking up, one idea we has was to let Codi rank them on priority. What do you think about that?

 Priority ranking sounds right, as long as the full list stays visible somewhere so nothing waits silently forever. My vote would be to surface whichever decision is blocking the most downstream work first.

Maker

 Definitely worth looking into and experimenting more. Thank you for your input!!

💎 Pixel perfection

The “I didn’t want a bigger cockpit, I wanted to hand the work to someone and step away” line really captures the problem. a lot of agent tools save time on coding, then quietly create a new job around choosing models, managing context, watching runs, and deciding what happens next. As someone building across multiple parts of a product, the idea of managing one AI CTO instead of babysitting a fleet of agents sounds much closer to the experience I’d actually want :)

The automatic model routing and overnight work are especially interesting, but also where trust matters most. Curious how Codi decides which decisions genuinely need the founder, and what safeguards stop it from confidently moving too far in the wrong direction while nobody is watching?

Maker

 You captured our thought process so well. There are an initial set of ruleset like problems including monetary and security - those are the ones escalated to human. Rest Codi answers as a supervisor. This mainly abstracts the technical decisions away.

The PRs are small enough to do no major harm :)

Maker

 Codi works against the project charter and feature map it builds with you at the start. That gives it a clear definition of what is in scope, what is out of scope, and what success looks like for each task.

The work also stays contained in isolated git worktrees and produces reviewable diffs. Tests, acceptance criteria, and a separate review pass have to succeed before Codi considers a task complete. If the work changes product scope, introduces a security or spending decision, or conflicts with the original charter, it comes back to you through the inbox.

We want unattended work to mean "useful progress waiting for you," not "surprise, your product changed overnight." Thanks for raising this because that trust boundary is one of the most important parts of what we’re building.

Congrats on launching a neat product! QQ: the cost story depends on Codi correctly calling a task 'boring' and handing it to a cheap model, and the tasks that look boring but aren't are exactly where cheap models produce plausible-wrong output. How are solving this conundrum if I didn't invent it at all and it exists?

Maker

 Not boring but suitable. Our algorithm uses public data to see which models perform well for which task. It is about reaping the cost benefit without loosing the accuracy too much. Plus the systems in place (skills and supervisor) boost the accuracy instead of dropping it. It is not a pure llm router, llm just classifies the task as to what kind of task it is and then we choose the best cheapest model for it. Hope that cleared it up a bit.

Maker

 To add to Shreyans’ answer: routing isn’t a one-way decision. The engineer’s output still has to satisfy the task’s acceptance criteria, tests, and review. If the result exposes unexpected complexity or fails validation, Codi can treat that as evidence that the task was misclassified and escalate it to a more capable model. The savings come from starting efficiently, not from being committed to the first model choice.

💡 Bright idea

Love the “manage one person, not a fleet of agents” framing! Does the inbox only surface blockers, or can I also review the decisions Codi made on its own?

Maker

 Currently inbox is for blockers but the feature map of each project stores the decision made by codi + every tasks details board also highlights them but they definitely also have a place in the inbox. Thank you for the feedback, definitely going up in the next update.

Maker

 this is a good feedback. We agree that important decisions deserve a dedicated view in the inbox too. We’re adding this to an upcoming update, thanks!

OpenAI
OpenAI Day
How did GPT-5.6 change the ambition or scope of what you shipped?
Maker
AskCodi is model-agnostic, so a jump like GPT-5.6 raises our ceiling directly, and the change was mostly about trust. Before, the models were good enough to demo an autonomous "AI CTO" but not reliable enough to let it run a team unattended, so we kept a human on every step and scoped the product down to stay safe. GPT-5.6 was the point we stopped hedging. Agents could take longer, more autonomous runs and actually come back with usable work, so we shipped the things we'd been too nervous to ship: Codi running in your off-hours, and running several projects in parallel. It also sharpened our core bet on cost. As the frontier gets both smarter and cheaper, routing the boring work to small models and saving the expensive calls for the hard parts saves people even more, so "a fraction of the cost" got stronger, not weaker.

Even if we clearly explain the full product vision and guidelines to Codi upfront, how do we ensure the AI agents it runs don't drift off course during execution, especially on longer, multi-step tasks? What mechanisms are in place to keep them aligned with the original intent throughout, not just at the start?

Maker

 The system has been tested to build out complete applications from start and we have seen certain behaviours from agents that is preventing it from drifting 

  1. Codi truly acts as a supervisor - reviewing not just the PRs but also the plan

  2. Agents are escalating to human users when some request doesn't align with the vision or user experience

  3. Daily briefings shows what is built and what is left at any given time. Purposefully built to always stay updated of what agent is working on and correct as fast as possible

Agents are problematic in that format but alignment is core of the product we have built.

How do balance cost savings without sacrificing output quality?
Maker

 We have a few things built in that improve output quality while reducing cost. Codi acts as both a supervisor and reviewer for the other agents, and it also routes each task to the most cost-effective model using our routing algorithm. The routing doesn't rely only on benchmark scores - it also considers recent model performance and task difficulty. So instead of defaulting to Fable or GPT-5.6 Sol for everything, it'll often choose models like GLM 5.2, Kimi K3, or MiniMax M3 when they're sufficient. Across our benchmark suite, we've seen roughly 50% lower inference costs with only about a 0.5% drop in task completion.

Automatic routing based on complexity is smart, but will have to try and see! Thanks for the detailed explanation

Congrats on the launch guys 🥳! Quick question though I'd usually want to run 2-3 branches on the same repo at once (say a feature, a bugfix, and a refactor in parallel). Does AskCodi isolate that at the branch/worktree level within a single project so they don't conflict, or is the parallelism mainly across separate projects?

Maker

 Thanks! 🙌 Great question. Isolation is at the worktree level: each task gets its own branch and git worktree, so a feature, a bug fix, and a refactor never share a working tree or step on each other. Right now Codi runs them isolated and coordinates them rather than firing all three at the exact same instant, and true simultaneous same-repo branches are on the near-term list. Across projects it's already fully parallel.

Congratulations for your launch. Does Codi work across multiple projects simultaneously, or is it limited to one repo at a time like other AI coding tools? The description mentions "parallel across all your projects", just wanting to confirm if that means true multi-project orchestration from a single inbox.

Maker

   Yes, it is true multi-project orchestration. You can run multiple Codi sessions across different projects simultaneously, while approvals and questions from every session arrive in one inbox. You do not need a separate terminal or agent manager for each repo.

Maker

 It is truly multi project. Codi has a /standup mode where it can starts sessions across multiple projects based on need and token usage left.

I have a small question for the dev team:

What kind of tasks has it been best at for you so far? Trying to figure out where it'd fit in my workflow.

Maker

 if you are working across multiple projects, understanding each projects needs and prioritising your token usage is best application we have seen. Coding is good but it outshines on orchestrating.

 I see. So have you guys done good with managing the tokens efficient to optimize the operation at first

1234
Next