Imbue vs Khoj (2026)

Sculptor works with local git repositories, keeping files and git state in sync between an agent's container and the developer's local repository through Pairing Mode, and works alongside the developer's own IDE and environment. Docker gives each agent a container with a full copy of the repository, and a later release added devcontainer port forwarding.

Linters including pylint, ruff and pyre run as part of the verification checks. Supported coding agents are Claude Code and Codex, and Imbue invites requests for others. There is no connection to business systems such as CRM, ticketing, messaging or data warehouses, and tool calls within a task are made by the underlying coding agents, not by Sculptor.

Sourceimbue.com sculptor product page and Sculptor announcement, public release notesread 2026-08-31

Agents can be given tools, including web search and a code execution sandbox. Document ingestion covers local files and folders synced through the desktop, Obsidian and Emacs clients, a Notion workspace synced from the web app, and direct upload. Supported formats are PDF, Markdown, org-mode, Word, Notion and images. Khoj names no third party application connectors for CRM, ticketing, messaging, warehouse or calendar systems, and no Model Context Protocol tool support.

Sourcedocs.khoj.dev setup sync-your-knowledge, and clients documentation, github.com/khoj-ai/khoj READMEread 2026-08-31

Five or more coding agents can run in parallel, each in an isolated worktree with its own branch, terminal and diff view, all visible and mergeable from one workspace, and skills run as full agents that can spawn parallel subagents. A developer reviews and merges each agent's changes, and Sculptor opens GitHub pull requests and tracks their status.

Sourceimbue.com/product/sculptor and github.com/imbue-ai/sculptor READMEread 2026-09-29

Users can create agents with custom knowledge, persona, chat model and tools to take on any role. They can schedule automations for repetitive research and run an autonomous deep research mode that plans, browses and synthesizes results. Intermediate steps such as intent detection and web search are model driven, and the Default and Advanced model fields in ServerChatSettings configure them separately. A code execution sandbox runs Python as part of a task. Khoj names no multi agent coordination, delegation between agents, conditional branching or subworkflows.

Sourcegithub.com/khoj-ai/khoj README, docs.khoj.dev setup and navigationread 2026-08-31

The CI Babysitter dispatches agents automatically to fix failing pipelines, so a CI failure starts work on its own, and Sculptor opens GitHub pull requests and tracks their status. Otherwise a developer launches agents on demand from the desktop app on macOS or Linux. The CI Babysitter rests on one line in the product page FAQ and is not in the README's feature summary.

Sourceimbue.com/product/sculptor and github.com/imbue-ai/sculptor READMEread 2026-09-29

Automations schedule repeated jobs, so Khoj runs a query automatically at a set time and frequency in the user's local time zone and emails the research to the user. Users configure, share and delete them from an automations page. Self hosted installs need Resend and authentication for the email. Clients cover web, desktop, Obsidian, Emacs and WhatsApp. Khoj names no inbound webhook or external event trigger.

Sourcedocs.khoj.dev features/automations and clientsread 2026-09-29

Each coding agent runs in a container with a full copy of the developer's repository, so it works from the whole codebase, not retrieved fragments, and devcontainer support supplies the project's environment configuration. Instruction audit rules that the developer writes in plain English persist as policy applied to agent work. Sculptor keeps no index, ingestion pipeline, document or wiki source, curated knowledge object or retrieval setup beyond the repository, which the customer maintains through git and which is copied into the container.

Sourceimbue.com sculptor product page and Sculptor announcementread 2026-08-31

The customer's own documents are indexed, including PDF, Markdown, org-mode, Word, Notion files and images, with semantic search over them and verifiable citations returned with answers. The desktop, Obsidian and Emacs clients keep local files and folders synced to the server, a Notion workspace syncs directly from the web app, and files can be uploaded directly in the web app. The index is held in the customer's own Postgres database on their self hosted instance. Local models through Ollama, vLLM or LMStudio can query it, so no data leaves the machine.

Sourcedocs.khoj.dev setup sync-your-knowledge and documentation, github.com/khoj-ai/khoj READMEread 2026-08-31

Sculptor has no memory between sessions. Neither the product page nor the README describes saving, resuming or forking a session. The launch announcement said every agent session was saved with its plans, chats, tool calls and code changes and could be resumed or forked from any point, but that no longer appears. Agents do not learn from corrections, carry conclusions between separate sessions or update their own instructions from experience, and they are kept apart from one another on purpose.

Sourcegithub.com/imbue-ai/sculptor README and imbue.com/product/sculptorread 2026-10-06

Agents are created with their own knowledge. The document index and conversation history persist in the customer's own Postgres database on their self hosted instance, and the desktop, Obsidian and Emacs clients keep files and folders continuously synced. State survives sessions, restarts and upgrades, and the customer owns it.

Khoj names no way for an agent to update its own knowledge or instructions from experience, carry conclusions between conversations or build up learned context over time. A pull request title in the project changelog mentions disabling memories for users, but Khoj does not describe a memories feature.

Sourcegithub.com/khoj-ai/khoj README and changelog, docs.khoj.dev setup sync-your-knowledge documentationread 2026-08-31

Each Sculptor agent works in its own worktree and branch and pushes to a local git remote, and its changes reach the codebase only when the developer reviews the diff and merges, which Imbue calls full review before merge. The merge is the customer's own gate, not an approval step inside the agent's run. Instruction audits, plain English rules checked against an agent's work, were listed as coming next in the Sculptor announcement but do not appear on the current product page or README.

Sourceimbue.com/product/sculptor, Sculptor announcement and github.com/imbue-ai/sculptor READMEread 2026-09-29

Agent code executes inside a sandboxed environment. Answers return verifiable citations, so a user can check sources before acting, and the assistant's search scope can be constrained to reduce fabrication. An administrative panel sits behind credentials created at first run. Anonymous mode, which bypasses login, is a deliberate opt in for local single user installs. Khoj names no approval gate, checkpoint, pause for human review, per tool authorization or queue of pending actions.

Sourcedocs.khoj.dev setup and privacy documentation, github.com/khoj-ai/khoj READMEread 2026-08-31

Each coding agent runs in its own Docker container, so it can install packages and run code without risking the developer's machine, unlike git worktrees, which share the local environment. Sculptor runs fully locally, with data sharing levels that control what goes back to Imbue.

The Essential only level shares just error logs and basic usage, which Imbue calls the safe choice for work code, and code is not used for training by default. There is no security certification, trust center, single sign on, role based access control or team audit capability, since the product is a desktop app for a single developer, in beta.

Sourceimbue.com sculptor product page and Sculptor announcementread 2026-08-31

An admin panel sits at the server admin path, with an administrator account and password created during first run. A separate path sets up authentication, and an explicit anonymous mode for local single user installs bypasses login. Deployment configuration covers allowed domains, HTTPS behavior and trusted origins for a reverse proxy.

By default Khoj is only reachable on the machine it runs on, and remote access needs deliberate configuration. The customer holds the database. Khoj names no security certification, attestation, single sign on or SCIM provisioning. In a self hosted open source deployment, the operator owns the perimeter.

Sourcedocs.khoj.dev setup, privacy and advanced authentication and remote documentationread 2026-08-31

Sculptor shows a diff for each agent's changes, which is review of output, not a record of each step the agent took. Pairing Mode brings an agent's work from its container into the local repository with files and git state in sync, so changes can be inspected in the developer's own editor and diff tools, and merges show conflicts to resolve.

At launch, every agent session was saved with its plans, chats, tool calls and code changes intact and could be resumed and forked from any point, but the current product page and README no longer describe this. Suggestions, which flagged misleading agent behavior such as tests that falsely report passing, no longer appears either. There is no analytics across sessions, fleet reporting or organizational dashboard.

Sourceimbue.com/product/sculptor and github.com/imbue-ai/sculptor READMEread 2026-10-06

Answers return verifiable citations to the source documents they were grounded in. An administrative panel at the server admin path lets the customer review configuration in their own database. The server writes log output, and the customer holds the Postgres database directly. Khoj names no agent execution trace, tool call record, run history view, audit log of user actions or analytics.

Sourcedocs.khoj.dev setup and privacy documentation, github.com/khoj-ai/khoj READMEread 2026-08-31

Sculptor runs on the customer's own computer, not Imbue's servers. Data never leaves the machine, Imbue does not store repositories or train on code, and Sculptor is free, local and open source under the MIT license. Each agent works in an isolated git worktree with its own branch and terminal, and a Docker or remote container backend is experimental. It runs on macOS on Apple Silicon and on Linux, including ARM64.

Sourceimbue.com/product/sculptor and github.com/imbue-ai/sculptor READMEread 2026-09-29

The product is AGPL-3.0 open source and self hosted. It installs through Docker, Docker Compose or pip on macOS, on Windows through WSL2, and on Linux, backed by the customer's own Postgres database or an embedded database option. Environment variables configure the domain, HTTPS, allowed hosts and reverse proxy settings.

Remote access is set up separately, and Khoj can start automatically through cron or Task Scheduler. Data never has to leave the customer's private network, and Khoj works without an internet connection when deployed on a personal computer. Khoj Cloud has been sunset, so self hosting is the only way to deploy.

Sourcedocs.khoj.dev setup, privacy and advanced remote documentation, app.khoj.dev sunset noticeread 2026-08-31

Sculptor ships a library of workflow skills that run as full agents with their own tools. They cover spec writing, mock generation and a fix bug skill that runs a reproduction interview, writes failing tests and then makes them pass, with experimental skills listed separately and a CI Babysitter that fixes failing pipelines. The set is small and each skill is named.

Sourcegithub.com/imbue-ai/sculptor README and imbue.com/product/sculptorread 2026-09-29

Users create each agent themselves, with custom knowledge, persona, chat model and tools, and a Khoj blog post gives a step by step guide to building custom agents. Khoj publishes no template library, preset or prebuilt agent set, shareable agent format, community gallery or starter pack. Khoj Cloud, which hosted any shared agent surface, has been sunset, so a new user begins from an empty self hosted installation.

Sourcegithub.com/khoj-ai/khoj README, docs.khoj.dev navigation, app.khoj.dev sunset noticeread 2026-08-31

Sculptor offers Claude models out of the box (Opus 4.8, 4.7 and 4.6, Sonnet 4.6, Haiku 4.5) on the customer's own Claude plan. It also supports the Pi agent harness, through which the customer can choose any provider or switch models mid session, and it can run any terminal based agent, so the choice spans model makers. The product page FAQ still calls Pi support experimental, and Codex is no longer listed.

Sourceimbue.com/product/sculptor and github.com/imbue-ai/sculptor READMEread 2026-09-29

Chat models are configured per provider through an AI Model Api object that holds the credential. OpenAI, Anthropic and Gemini are supported, alongside offline and local models through any OpenAI-compatible server, including Ollama, vLLM, LMStudio and llama-cpp-server, by setting the API base URL.

Each chat model is registered with optional tokenizer and max-prompt-size fields and its own vision-enabled flag, and users select their preferred model in their own settings. ServerChatSettings holds separate Default and Advanced model choices for intermediate steps such as intent detection and web search. Agents can be created with their own chat model.

Sourcedocs.khoj.dev setup add-chat-models and advanced ollama and lmstudio documentation, github.com/khoj-ai/khoj READMEread 2026-08-31

Customers can build their own Sculptor plugins in JavaScript or TypeScript and swap the agent harness entirely. Sculptor supports the Pi harness alongside Claude Code and can run and manage any terminal based agent. It is open source under the MIT license, with a public repository and documentation. The plugin interface is described on the product page, but no plugin reference has been recorded.

Sourceimbue.com/product/sculptor and github.com/imbue-ai/sculptor READMEread 2026-09-29

The server, clients and docs are open source under AGPL-3.0 in one public repository. The desktop, Obsidian, Emacs and WhatsApp clients talk to a self hosted server over HTTP at a configurable URL, so customers can read, modify and run the product themselves. There is no API reference, SDK or MCP server, so a customer's own code has no published interface for invoking Khoj agents or automations.

Sourcedocs.khoj.dev sitemap and github.com/khoj-ai/khojread 2026-09-29

The Sculptor announcement listed Suggestions as a beta that points out fixes to agent code, and listed instruction audits and flagging misleading behavior, such as tests that pass without real tests, as coming next, so those were roadmap features.

The current product page and README describe diff review for each agent, a fix bug skill that writes failing tests before fixing, and a CI Babysitter, and mention neither Suggestions nor audits. Tests the agent writes and runs are part of its method, not a harness scoring its output. Imbue's separate open source tool for catching agent mistakes is another product.

Sourceimbue.com Sculptor announcement, imbue.com/product/sculptor and github.com/imbue-ai/sculptor READMEread 2026-09-29

Verifiable citations let a human reader check an answer's sources after the fact. Khoj names no testing, evaluation, scoring, benchmarking or regression capability, and no self check or verification gate on agent output.

Sourcegithub.com/khoj-ai/khoj README, docs.khoj.dev navigation and setup documentationread 2026-08-31

Coding agents run code, install dependencies and run terminals and test suites inside isolated Docker containers, which is code execution, not operation of software without a programmatic interface. Devcontainer port forwarding lets a developer view an app running inside a container from their own browser, so the person, not the agent, looks at the output. All agent actions reach the filesystem, git, package managers and linters programmatically inside the container. No agent controls a browser, navigates, fills forms, interacts with a screen or automates a desktop.

Sourceimbue.com sculptor product page and Sculptor announcement, public release notesread 2026-08-31

Agents run Python in a code execution sandbox and reach the web through retrieval and semantic search over fetched content. Khoj names no browser control, navigation, form filling or screen interaction. The web, desktop, Obsidian, Emacs and WhatsApp clients are ways to reach Khoj, not interfaces Khoj operates.

Sourcedocs.khoj.dev setup and clients documentation, github.com/khoj-ai/khoj READMEread 2026-08-31