AGENT PROFILES

CodeWhale

Shannon Labs · Coding agents

Open-source coding agent for the terminal, formerly DeepSeek-TUI, that works with many model providers and local models.

Typical tasks

Availability in Europe
Added to the catalog on Open-source coding agent for the terminal, formerly DeepSeek-TUI, that works with many model providers and local models.Details
01

What it does

7 features

CodeWhale · Open-source terminal coding agent

Product and access

The project was first called deepseek-tui and has shipped as Codewhale since v0.8.41 (23 May 2026). The repository now lives under codewhale-hq, and the old paths under Hmbown redirect there. According to the README it is independent and not affiliated with DeepSeek or other model providers, although DeepSeek is the default provider. The catalog lists DeepSeek Harness separately.

Core features

7 features from provider sources · checked on 5 November 2026. Availability and limits are directly below.

  • Tasks in the terminal

    You give a task, and Codewhale reads the repository, edits files and runs commands until the task is done or it needs your input.

    Source 12
  • Many models, also local

    More than 40 providers are built in, with deepseek-flash as the default. `/model` switches between turns, and `/model auto` picks a model per turn.

    Source 12
  • Plan, Work and Operate modes

    Plan only reads and changes nothing. Work edits files and runs commands. Operate has the same permissions and works toward a goal until it is verified or needs your input.

    Source 13
  • Receipts and undo

    `/receipts` lists changed files, commands, web and MCP calls, approvals and failures. `/undo` takes back the last request including its file changes.

    Source 14
  • Subagents, Fleet and workflows

    Tasks can be split across roles such as explore, implement or reviewer. Workflows add ordering, result handoffs and checkpoints.

    Source 15
  • Pull request reviews

    `codewhale review --pr` creates an advisory review. The GitHub bot publishes findings with separate jobs and, according to the docs, does not run code from the pull request.

    Source 16
Show all 7 features
  • Permissions and sandbox

    Levels Ask, Auto-Review and Full Access. In Ask and Auto-Review, commands run in an OS sandbox where available: Seatbelt on macOS, bubblewrap on Linux when installed.

    Source 17

Usable AI models

Open choice

More than 40 providers are built in, plus OpenRouter, and any OpenAI-compatible endpoint works. Locally CodeWhale runs with Ollama, vLLM or SGLang, without a key. You pick the model yourself or let /model auto decide for each turn.

Model families named

  • Anthropic (Claude)
  • OpenAI (GPT)
  • DeepSeek
  • Google (Gemini)
  • Mistral
  • Moonshot (Kimi)
  • xAI (Grok)
  • Other models can be connected freely
Model choice
Automatic or your choice
Your own API key
Yes
Router or custom endpoint
Yes
Local models
Yes

The rating covers the generally available main access. Documented · Sources accessed on

CodeWhale: README

02

Does it fit my task?

2 practical tasks · 4 requirements and limits

A good fit if …

  • you want an open-source terminal agent without a required account
  • you want to choose providers freely or use local models
  • you want to trace what a run changed and executed

Less suitable if …

  • you need a managed platform with a contract, roles and data location commitments
  • you rely on an OS sandbox on Windows
  • you do not want to deal with defaults for telemetry and approvals

Editorial assessment based on provider documentation. Not a product test.

Practical tasks

Editorial examples based on documented features. No product tests performed.

01

Fix tests

Your input
A clean Git state and the task of fixing failing tests.
Possible result
Codewhale edits files, runs commands after approval and explains the change.
What to check
Check the diff and `/receipts`, and use `/undo` if needed.
02

Run a task in CI

Your input
`codewhale exec --auto --sandbox workspace-write` with your own model key.
Possible result
A task runs without an interface and can be traced afterwards through receipts.
What to check
Set the sandbox explicitly and evaluate `codewhale receipts --last`.
EDITORIAL EXAMPLE

Have it fix failing tests and explain the change.

Your possible prompt

Read the project rules. Investigate [error], propose a small change, and run the relevant checks. Show me the diff and any remaining uncertainties.

Adjust the task, data and permissions.

Requirements and limits

  • Requirement

    Model access required

    According to the install guide, Codewhale does nothing useful without a model. You need a provider key or a local model.

    Source 7
  • Limitation

    Maturity before version 1.0

    The current version is v0.10.2, marked as a regular release. Security patches only go to the latest stable version. The change list is long and frequent; several parts are labeled experimental or preview.

    Source 18
  • Limitation

    Mind the default behavior

    In Ask, file edits in the repository run without asking. `exec --auto` uses Full Access without a sandbox by default unless you set `--sandbox workspace-write`. Windows has no OS sandbox yet.

    Source 12
  • Not yet confirmed

    Status of the hosted offering

    The website lists a hosted web app as in development, while the guide describes its phased shutdown since September 2026. The privacy policy covers accounts and managed models. How far this is usable today remains open.

    Source 19
03

Can I use it?

With requirements

Europe & regional availability

Read the README, the telemetry docs and the privacy policy (as of 4 October 2026). The policy has no section on the EU, the EEA, the UK or Switzerland.

Reviewed scope: Local program; hosted accounts reviewed separately
Reviewed on:

  • With requirements

    Data path with your own key

    Request content goes to the provider you connect, under that provider's terms. The default provider is DeepSeek. Local models through Ollama, vLLM or SGLang need no provider key.

    Region, reason and source

    Region covered: EU/EEA, United Kingdom and Switzerland

    Documented reason: Location and processing depend on the chosen provider; Codewhale makes no commitment on this.

    Official source · codewhale.net · Official source · github.com

  • Availability still unclear

    Accounts and location

    The policy describes one global product without a residency selector. Sign-in and session data live in Cloudflare Durable Objects, currently configured in the US jurisdiction. The local program needs no account.

    Region, reason and source

    Region covered: EU/EEA, United Kingdom and Switzerland

    Documented reason: The policy makes no commitments for European users.

    Official source · codewhale.net

  • With requirements

    Usage counting

    Counting is on by default and sends aggregate values with a random ID that rotates every 90 days. Content, code and credentials are excluded. PostHog may act as a processor. `codewhale config set telemetry false` turns it off.

    Region, reason and source

    Region covered: EU/EEA, United Kingdom and Switzerland

    Documented reason: Can be turned off; the docs name no European location.

    Official source · github.com · Official source · codewhale.net

Europe and regulation explained

Accounts & permissions

Based on provider documentation. Permissions depend on the account, integration and organization policies. Review the permissions actually requested before use.

Configuration covered: Local program without a required account

Sources reviewed:

Show all 3 points
  • Documented

    Account used

    According to the README you need no account. Keys from `/provider` stay on your computer. Experimentally, a ChatGPT plan can be used instead of an API key. The README describes no roles or SSO for the local program.

    Evidence
  • Documented

    Admin approval

    Ask prompts before commands that are not read-only; `y` once, `a` for similar commands in the conversation, `n` refuses. Auto-Review lets rules and a reviewer model approve routine calls. Full Access never asks.

    Evidence · Evidence
  • Documented with conditions

    Data access & team sharing

    Inside the sandbox, commands cannot read credentials such as `~/.ssh`, `~/.aws` or `.env`. Full Access turns the sandbox off, and Windows has none yet. Subagents get narrowed permissions.

    Evidence · Evidence
04

Memory and conversation continuity

4 of 6 criteria documented

Markdown files with SQLite search, global or per repository; can be turned on, offline and managed by command.

Local memory, off by default

Sources accessed:

Documentation by criterion

StorageDocumented

Storage and scope

Memory is off by default. Once on, it lives under `~/.codewhale/memory/` as Markdown with a SQLite search index. It does not sync to the cloud, but entries go to the chosen provider as prompt context.

SelectionDocumented

Selection and timing

You add with `#` in the input box or `/memory native remember`. The model can use the `remember` tool without approval. According to the docs, secrets and transient task state do not belong there.

Edit & deleteDocumented

Inspect, edit and delete

`/memory native` can search, read, export, reindex and delete, including everything at once. The files are plain text and can be edited directly.

Access boundariesPartly documented

Agent and channel boundaries

The scopes are global and workspace; the workspace ID is a hash of the Git origin, and according to the docs one repository's notes do not appear in another. The page does not describe boundaries between agents or channels.

RetrievalDocumented

Retrieval

Up to 32 entries or 12,000 characters enter the system prompt as a marked, untrusted block. Beyond that, the model searches with `memory_search` and `memory_get`.

New conversationsPartly documented

Session reset and compaction

The memory state at session start belongs to the fixed prompt prefix, and later changes are appended to history. `/compact` summarizes long context. That memory outlasts this follows from file storage; the page names no separate reset rule.

05

Cost and effort

2 plans · Prices as of 5 November 2026

Prices and included usage

Pricing checked

The program is free. According to the website you need no account for the terminal, local web interface and GUI. Costs arise with the chosen model provider.

PlanPriceIncluded
CodewhaleFree under the MIT license

Terminal, local web interface, headless and Runtime API; model use not included.

Your own provider keyAccording to the provider's price list

`/cost` shows session costs when the price is known.

Price source

  • The website and README promote an OrcaRouter offer; according to the README, Codewhale earns a referral fee from it.
  • Third-party pricing was not used as evidence.

Pricing details

Setup

Install with the install script, winget, npm (Node.js 18 or newer) or Cargo (Rust 1.89 or newer), start in a Git repository and pick a key or a local model through `/provider`. The script needs curl or wget and a SHA-256 tool, but neither Node nor Rust.

Operation & oversight

Based on public provider documentation. Assessments are editorial inferences from the stated properties. Actual time requirements and reliability have not been measured.

Deployment covered: Local process; model calls go to the chosen provider

Sources reviewed:

Editorial assessment

Codewhale shows what it did and can take changes back. The defaults need attention: file edits without asking, counting turned on, Full Access with `exec --auto`. The project is before version 1.0 and releases at short intervals.

Show all 6 points
  • Documented

    Files and working state

    New installs write to `~/.codewhale/`; old data from `~/.deepseek/` stays readable. Sessions are saved, and `/restore` rolls the workspace back to an earlier snapshot.

    Evidence · Evidence · Evidence
  • Documented

    Sign-ins

    You enter keys through `/provider`, and they stay local. The sandbox blocks credentials for commands. `/preview-request` shows a redacted preview of the next request without sending it.

    Evidence
  • Documented with conditions

    Starting further runs

    Operate offers recurring work as an automation, created after your approval. Automations cannot hold more permissions than the session that created them. GitHub workflows start reviews on events. The README does not name which schedules are possible.

    Evidence · Evidence · Evidence
  • Documented

    Human intervention

    Approvals depend on the level. Plan refuses changes centrally and asks at the end how to continue. Every permission change appears in the transcript.

    Evidence
  • Documented

    Detecting errors

    `/receipts` lists failures and who approved what. `/turn inspect` shows a turn's model, cost and duration. `codewhale doctor` checks the installation.

    Evidence · Evidence · Evidence
  • Documented

    Resuming after failures

    According to the guide, sessions can be resumed and forked, and `exec --continue` continues headless runs. `/undo` and `/restore` take back changes; `/restore` overwrites later edits and needs trust mode or Full Access.

    Evidence · Evidence
06

Evidence and open questions

21 sources · 2 open points

Open points

The reviewed sources contain no reliable information on these points. A gap says nothing about the product.

Europe: Availability still unclear

Requirements and limits: not yet confirmed

Sources and evidence

Read the repository (README, docs, changelog, release list via the GitHub API), the website and the privacy policy; checked the redirect of the old repository path over HTTP. The project mostly writes Codewhale, sometimes CodeWhale. Feature claims come from the project; example tasks are editorial. Our own hands-on tests are pending.

Editorial team: Michael Meierhoff · sources reviewed on . Provider information is not independent evidence of performance.

Read profile as Markdown

How we assess the sources

Provider

The privacy policy calls Codewhale a product of Shannon Labs, Inc. The README names Hunter as maintainer, and the project accepts community contributions.

Operator
Shannon Labs
Organization
Project and operator
Provider sourcesSource 1 Source 2 Reviewed
Changes to this profile 1
  • Open-source coding agent for the terminal, formerly DeepSeek-TUI, that works with many model providers and local models.

These notes describe editorial changes. Source and pricing review dates appear in their respective sections.