Skip to content
KOLDOS
Sign in
KOLDOS

What KOLDOS can do, and what it can't yet

Each capability is described by what it does, what it's for, its real status, where it runs and what you need to use it. Nothing here is marked available unless it works in the current build.

Chat

Available
What it does
Talk to KOLDOS in a conversation. Every message is a turn: KOLDOS decides a route, gathers context and answers with the model on your machine.
What it's for
Questions, explanations and clarifications about your code and your project.
Current status
Available. Works in the current build.
Where it runs
On your computerThe model runs on your own hardware through llama.cpp. KOLDOS sends no telemetry.
What it requires
  • KOLDOS installed
  • The KOLDOS 7B model file
Related
KOLDOS 7BRoadmap

Coding

Available
What it does
KOLDOS is a programming assistant first. It reads, writes and edits files in your project, runs commands and checks the result.
What it's for
Fixing bugs, writing features, refactoring and answering questions about a codebase.
Current status
Available. Works in the current build.
Where it runs
On your computerFiles are read and written on your disk, inside the workspace folder only.
What it requires
  • A project folder opened as the KOLDOS workspace
  • Everything chat needs
Related
KOLDOS 7BRoadmap

Local execution

Available
What it does
The model, the tools and the memory all run on your computer. There is no cloud mode.
What it's for
Working with no per-request cost and no prompts sent to a model server.
Current status
Available. Works in the current build.
Where it runs
On your computerInference, tools and memory run on your hardware.
What it requires
  • An NVIDIA GPU with CUDA, measured on 6 and 8 GB cards
  • Enough system RAM for the rest of the model and the app
Related
KOLDOS 7BRoadmap

Tools

Available
What it does
KOLDOS calls tools inside a bounded loop: read and write files, search the project, run shell commands. 25 tools exist and only the ones a turn needs are exposed to the model.
What it's for
Acting on the project instead of guessing, and checking results against the real files.
Current status
Available. Works in the current build.
Where it runs
On your computerEvery tool call goes through the Tool Manager, the only boundary to disk and network, and is limited to the workspace by a capability token.
What it requires
  • A route that allows tools, chosen per turn by the Decision Engine
Related
KOLDOS 7BRoadmap

Files

Available
What it does
Attach files to a turn and browse the project tree. Attachments are copied into the workspace so the turn reads what you attached.
What it's for
Working with logs, documents and code that is not in the project yet.
Current status
Available. Works in the current build.
Where it runs
On your computerAttachments are copied to a folder inside the workspace on your disk.
What it requires
  • Enough context for the parts of the file the turn needs
Related
Roadmap

Local API

Available
What it does
A REST API with streaming on your own machine. Any client that speaks the common chat completions format can use the whole KOLDOS turn, not just the model.
What it's for
Editor extensions, scripts and bots running on the same computer.
Current status
Available. Works in the current build.
Where it runs
On your computerListens on 127.0.0.1 by default. Requests never leave your computer.
What it requires
  • Running koldos serve
  • An optional access token for everything except the health check
Related
Roadmap

Memory

Experimental
What it does
KOLDOS remembers project rules, preferences and past corrections in a local database, isolated per project. Unused memories lose weight over time but are never deleted.
What it's for
Not repeating the same rules, and warning before a mistake that already happened once.
Current status
Experimental. Usable today, but behavior and interfaces can still change.
Where it runs
On your computerStored in a SQLite database on your disk, one per project.
What it requires
  • Real use to calibrate thresholds, which are marked as not calibrated yet
Related
Roadmap

Vision

Experimental
What it does
Drop an image into the conversation and KOLDOS reads it: text through OCR, diagrams and charts through a small separate vision model.
What it's for
Reading error screenshots, diagrams and charts during a coding task.
Current status
Experimental. Usable today, but behavior and interfaces can still change.
Where it runs
On your computerOCR and the vision model run on the CPU of your computer. The image becomes text before it reaches the main model.
What it requires
  • The optional vision model files for images with little text
  • About 3.4 GB of free RAM while the vision model is loaded
Related
Roadmap

Hosted API

Planned
What it does
KOLDOS as a service: the same turn, run on hosted GPUs and reached with an API key.
What it's for
Using KOLDOS from apps, servers and Discord bots without local hardware.
Current status
Planned. Committed to, but work has not started.
Where it runs
Outside your computerRequests would be processed on hosted infrastructure. The design exists; the service has not started.
What it requires
  • Manual approval: contact itzkoldkir on Discord
  • An API key created in the dashboard
  • Credits on your account
Related
RoadmapDeveloper portalWhere the data goes

Automation

Planned
What it does
Acting on systems outside the project: a real browser, databases and external APIs, in manual, plan or automatic mode.
What it's for
Tasks that need to touch live systems, with permissions granted per domain and per plan.
Current status
Planned. Committed to, but work has not started.
Where it runs
Depends on the taskDepends on the system being driven. Everything that comes back from outside is treated as data, never as instructions.
What it requires
  • Permission broker and confirmations, designed and not implemented
Related
RoadmapWhere the data goes

Agents

Planned
What it does
KOLDOS is one assistant. It does not run independent agents; multi-step work is split into internal sub-executions with their own context and budget.
What it's for
Large tasks divided among planned sub-executions, with testers and reviewers.
Current status
Planned. Committed to, but work has not started.
Where it runs
On your computerSub-executions run inside the same local process, one level deep at most.
What it requires
  • A stable executor and planner
  • Planner teams, designed and not implemented
Related
Roadmap

In research

What is being planned

None of this is available today. It is planned or being researched, and may change.

Thinking V2

Planned

A planned, experimental reasoning system. Not available today. The goal is not to generate more thinking tokens, but a structured process that plans, researches, verifies, criticizes and rethinks before answering.

Questions it is meant to work through

  1. 01What am I trying to solve?
  2. 02What do I already know?
  3. 03What information is missing?
  4. 04Which parts are uncertain?
  5. 05What assumptions am I making?
  6. 06What could be wrong?
  7. 07What evidence do I need?
  8. 08What should I search for?
  9. 09Which tools can help?
  10. 10What hypotheses exist?
  11. 11Which alternatives should I consider?
  12. 12What contradictions exist?
  13. 13What should I verify before answering?
  14. 14Which part of my solution needs rethinking?
  15. 15Does the result really answer the original request?
  16. 16What errors could my solution contain?
  17. 17Do I need to run a test to check it?

Planned tools

Planned

  • Search Web

    Web search as a tool the model can call.

  • Deep Web

    Deeper searches across more sources than a single web search.

  • Terminal Test

    Run and check commands or code in a controlled environment.

  • Web Tools

    Tools similar to web developer tools, to get information a conventional search can't provide.