Our npm install of Odysseus at commit 9fdeb64 succeeded in 3 seconds. The separate Docker attempt stopped at a nested-mount safety check. Odysseus suits local-model operators who want to experiment with a combined workspace and can trace its connections. Readers whose daily work requires dependable MCP execution should choose another route: issue #1789 reports connected integrations that chat and agents could not call.
Chat needs a model connection
Connect an existing local or API model endpoint for chat; Cookbook provides a separate model-serving workflow. The Odysseus README lists chat, agents, research, documents, email, notes and calendar alongside local-model tools. Its feature inventory also includes shell and file tools, skills, memory and MCP integrations. Compare presents model answers side by side, while Cookbook handles model recommendations, downloads and serving.
W J Pearce's walkthrough makes the connection concrete. The dashboard initially showed "No models connected"; Pearce then connected an existing Ollama server and obtained a greeting from gemma4:26b. Ollama handled inference in that arrangement. The official setup guide documents host.docker.internal:11434/v1 for an Ollama server on the Docker host, and explicitly distinguishes that connection from Cookbook Serve.
What people use odysseus for
W J Pearce reports a completed local greeting, while GitHub users describe model-serving and integration tasks with specific failures and recoveries. A greeting exercises the model connection. An Axonius query also requires integration discovery, tool selection and execution. These examples identify who attempted each task and what they reported, so a launched research request is distinguishable from a completed conversation.
| Person and task | Reported outcome | What the example shows |
|---|---|---|
| W J Pearce: local chat through Docker and Ollama | gemma4:26b answered a greeting on the author's hardware | A completed conversation using an external inference server |
| W J Pearce: research an NVIDIA Spark and CPU announcement | Started a Deep Research request | The author describes launching that particular research workflow |
| Lemmy commenter quoted by The AI Future: compare local models | Described multiple answers and voting after connecting Ollama | A secondhand account of model comparison |
| nowweh: serve Qwen2.5-Coder-7B GGUF on a GTX 1070 Ti | Reported success after using a precompiled CUDA llama.cpp build and matching libraries | A machine-specific recovery after CPU fallback |
| updateset: query Axonius for Linux devices with port 8080 and public IPs | Reported restored MCP and email access after repairing ChromaDB | One reporter's recovery from connected but uncallable tools |
Pearce's greeting is a useful first acceptance task because it checks the endpoint and response path together. The Axonius request asks the model to reach a named integration and submit a particular query. A working chat connection and a completed agent action therefore answer different installation questions, even when both begin in the same browser interface and use the same model endpoint.
What happened when we ran it
Our run installed the npm dependencies in 3 seconds at commit 9fdeb64 on November 19, 2026. It recorded 5 dependencies and 62 MB on disk, with 0 known vulnerabilities from npm audit. The repository scan found 3,322 files, about 852,582 source lines and a 56 MB checkout. The harness recorded 3 CPUs, 8 GB RAM and image label lab-node:22.
The separate Docker attempt used the project's Compose configuration with every port bound to 127.0.0.1. Its safety step rejected two overlapping mount pairs: data with data/ssh, and data with data/huggingface. The recorded safety rule was "mount sources may not be nested". This was a lab-policy refusal at the mount check; the teardown record reports successful cleanup.
| Question | Answer | Verdict | Why |
|---|---|---|---|
| Did npm installation complete? | Succeeded in 3 seconds; 5 dependencies recorded; 62 MB on disk | Completed | Node dependency installation finished |
| Did the harness build the application? | No build script/target; skipped | Skipped | The harness invoked no application build |
| Did the harness run tests? | No tests script/target; skipped | Skipped | The harness invoked no test suite |
| What did npm audit report? | 0 known vulnerabilities: 0 critical, high, moderate or low | Clean audit result | The result applies to the audited npm dependency set |
| How large was the checkout? | 3,322 files; about 852,582 source lines; 56 MB checked out | Measured | Repository scan at 9fdeb64 |
| Which project files were present? | 12 CI workflow files; Dockerfile, Compose file and tests directory | Present | Repository signals found by the harness |
| Did the Docker safety step pass? | Failed: nested data and data/ssh mounts; nested data and data/huggingface mounts | Blocked by lab policy | Mount sources may not be nested |
| Did teardown complete? | OK | Completed | The hands-on teardown record reports success |
The tests directory and skipped test step describe different things: repository contents and harness execution. GitHub identifies Python as the primary language, while the measured dependency command used npm. The 3-second result is dependency installation time, not application startup time. Build and test steps were skipped rather than passed, and the exact measured checkout remains 9fdeb64. The Docker refusal describes the lab's mount policy rather than an application error.
Where people get stuck: GPU detection and tool execution
Issue #831 reports visible NVIDIA GPUs alongside CPU-only Cookbook serving; issue #1789 reports connected MCP integrations unavailable to chat and agents. Both involve a connection indicator succeeding before the requested work succeeds. Their logs point to different stages: inference compilation and runtime setup in the GPU thread, and tool indexing, schema delivery or execution in the MCP discussion.
| Symptom | Reported finding | Diagnostic direction |
|---|---|---|
| NVIDIA GPU visible, model on CPU | Issue #831 records missing nvcc and CUDA Toolkit errors despite working nvidia-smi | Separate device passthrough from the inference build |
| GTX 1070 Ti with newer CUDA tooling | The thread identifies Pascal sm_61 compatibility problems with CUDA 13 | Match compiler support to the card |
| CUDA compilation produces cospi errors | The thread discusses CUDA-header and glibc compatibility | Read compiler and header errors separately from GPU detection |
| MCP connected but unavailable | updateset restored access after fixing ChromaDB | Inspect tool-index and ChromaDB logs |
| MCP unavailable with ChromaDB running | sombradeluna-tec reported tools_sent=0 for Obsidian | Inspect the schemas sent to the model |
| Native Windows shell calls fail | joebog1 described Linux assumptions in bash tooling | Trace the tool's platform assumptions |
nowweh's original GPU report names Qwen2.5-Coder-7B-Instruct-GGUF and a GTX 1070 Ti. The container could run nvidia-smi, but serving fell back to CPU. The build log reported missing nvcc and "CUDA Toolkit not found." Later discussion covers older-card compiler support and precompiled serving binaries. The thread describes diagnostic improvements as "diagnosis-only, no auto-install," separating detection work from changes to the serving environment.
The MCP reports diverge after the initial symptom. updateset said repairing ChromaDB on port 8100 restored MCP and email access. elpideus replied that ChromaDB was already running and the failure persisted. sombradeluna-tec reported Obsidian showing "Connected (12/12 tools)" alongside tools_sent=0, then later described successful integration after code and wrapper changes. The thread records individual failures and recoveries with different paths through the application.
Agent access needs a private trust boundary
The README says to keep AUTH_ENABLED=true for network access and LOCALHOST_BYPASS=false outside local development. Those settings accompany tools that can read files, run shell commands and use private service credentials. The official setup guide calls Odysseus an admin console, describes restrictions for non-admin users and tells operators to keep raw model and service ports internal.
The August independent review reports no shell or filesystem sandbox. Login, account privileges and action approval address different access decisions: identity, available functions and individual actions. For a deployment holding email or internal documents, the account's privileges and the host's accessible data determine what those tools can reach. The setup guide reserves shell/Python/file read/write access from non-admin users by default and identifies MCP management as admin-gated.
The Windows ZIP discussion records a separate security event. The reporter encountered a "Virus detected" download block and later said they had cloned the repository in issue #554. Commenters suggested explanations and another acquisition route. The reported clone answers how the user obtained the source; it does not explain the original warning or justify disabling download protection.
Maintenance work is visible, and contributors still collide
The November 11 update on test-hardening tracker #2523 says the calendar helper extraction and endpoint-resolver, provider-endpoint and service-health splits merged. Report-only focused-test guidance was implemented in CI. The remaining actionable work is regenerating tests/OVERSIZED_TEST_SPLIT_PLAN.md against the current test tree, superseding older tracker rows that still describe several slices as awaiting review.
The maintainer's closure criteria also require removing unintended cross-test helper imports, passing focused taxonomy and runner tests, completing whole-suite collection and refreshing the plan. Earlier comments document shared import-state helpers and clean-worktree fixes. High-risk route/API, database/session and security modernization belongs to the deferred Phase 2 scope. These dated updates provide a more specific maintenance picture than the number of workflow files alone.
Contributor coordination has a concrete failure example in issue #3629: dewanggaabdullah says document_tools.py work collided with a duplicate merge, then claims another file already covered by a different pull request. joebog1 uses ticked boxes for claimed files, while another comment uses checkmarks for completed or ongoing work. An exact filename, written status and pull-request link communicate which work is available more clearly than a box alone.
Odysseus vs Open WebUI, AnythingLLM and Claude
The independent reviews recommend Open WebUI for chat and AnythingLLM for document workflows. The AI Future also suggests Claude for readers who prefer hosted model access. Those recommendations divide the decision by task: a combined personal workspace, chat, documents or hosted assistance. They are the reviewers' recommendations, and the comparison below preserves that attribution rather than assigning performance or compatibility scores.
| Option | Cited reason to consider it | Decision boundary |
|---|---|---|
| Odysseus | Its README lists chat, research, documents, email and local-model workflows | Explore it when the combined workspace is the attraction |
| Open WebUI | odysseusai.dev recommends it for chat | Consider it when chat is the main job |
| AnythingLLM | odysseusai.dev recommends it for document workflows | Consider it when documents are the main job |
| Claude | The AI Future recommends it for readers who prefer paying for model access to managing hardware | Consider it when a hosted assistant fits the requirement |
There is a narrower interoperability comparison in issue #1789: updateset reports that the same MCP servers worked in OpenCode and Cursor while failing through Odysseus. That report concerns the user's existing servers and specific attempted requests. For an integration-dependent choice, the useful outcome is completion of that request through the application; the connected indicator was already present in the reported Odysseus failure.
Outdated claims, corrected
The current README identifies dev as the default branch and main as the curated alternative, correcting Cyber Notes' claim that an unqualified clone lands on main. The port discrepancy is different: The AI Future's localhost:7011 agrees with the README checked on November 19, while the detailed setup guide gives localhost:7000. The documents disagree, so the checkout's port mapping determines the address for that installation.
The independent setup page's claim that Cookbook's hardware scan gives an exact answer about what a machine can run overstates detection. Issue #831 reports visible NVIDIA hardware alongside CPU-only serving. The official setup guide separates passthrough from a CUDA-enabled llama.cpp build and names cudart discovery, toolkit availability and CPU layer placement as diagnostic signals. A detected device answers only the hardware-visibility part of that workflow.
The November 11 maintenance comment corrects the older tracker status: the named splits merged and report-only CI guidance was implemented. Plan regeneration and the closure checks remain part of that update. GitHub metadata records a November 18 push, while the latest-release lookup returned no release data. The email-authentication disagreement between the independent review and official setup guide is addressed with the unresolved implementation questions below.
The repository identifies this Odysseus
This review concerns odysseus-dev/odysseus, the self-hosted workspace linked by the current README and setup guide. That repository address distinguishes it from other software and services with the same name. Cyber Notes uses the former pewdiepie-archdaemon address in its walkthrough. For installation and credential configuration, the repository identity and current setup documentation provide the relevant connection between the named project and the software being operated.
Questions people ask
The README installs Odysseus through Docker Compose and prints the first admin password in container logs. Installation and startup appear in search questions; GPU fallback and unavailable MCP tools appear in GitHub reports. The answers below separate connecting an existing model server, serving a model through Cookbook and invoking tools through chat. Each path has its own documented commands or reported failure symptoms.
How do I install Odysseus AI?
The README's Docker instructions say to clone the repository, enter the odysseus directory, copy .env.example to .env and run docker compose up -d --build. The default branch is dev; main is the curated alternative. The setup guide says Compose starts Odysseus, ChromaDB, SearXNG and ntfy, and directs users to configure models, search and email in Settings.
How do I set up, run and start Odysseus?
The first admin password appears in docker compose logs odysseus, according to the README. After login, configure the model connection in Settings. For an Ollama server on the Docker host, the setup guide documents http://host.docker.internal:11434/v1. This connects to an already-running host server; Cookbook Serve handles the separate workflow of serving downloaded models through Odysseus and llama.cpp.
Why does Cookbook use the CPU when Docker can see my NVIDIA GPU?
Issue #831 records missing CUDA tooling after GPU visibility succeeds. Its original log says "Could not find nvcc" and "CUDA Toolkit not found," followed by CPU placement. The official setup guide distinguishes Docker GPU access from the CUDA-enabled inference build. The reporter's GTX 1070 Ti also introduces the older Pascal architecture compatibility discussed later in the thread.
Why are connected MCP tools unavailable in chat and Agent Mode?
Issue #1789 records several failure paths after integration discovery. updateset restored access after repairing ChromaDB; other users reported unavailable tools with ChromaDB running, including Obsidian output showing tools_sent=0. A contributor requests the exact server name, tool name, prompt and logs around the call. Those details locate the failure among schema delivery, tool selection, name resolution and execution.
Why does downloading the Odysseus ZIP trigger a virus warning?
Issue #554 records a blocked ZIP download and the reporter's later successful clone. Commenters suggest a possible false positive, but that explanation is their interpretation. The recorded event is a download blocked with "Virus detected." The README's installation route uses git clone, and the reporter's later message concerns obtaining the source through that route rather than identifying the warning's cause.
How can contributors understand the frontend, backend and state-management boundaries?
Issue #68 contains questions about frontend, backend and state boundaries, including confusion about the static directory and plain JavaScript. Those questions explain the contribution concern raised in the discussion. Issue #2917 gives a more focused example: execute_tool_block combines dispatch with individual tool implementations, and the proposed handler registry separates those responsibilities while retaining shared execution policy in the coordinator.
Can I contribute a small batch of tool refactors without overlapping another contributor's work?
Issue #3629 records agreement on smaller batches and a collision involving document_tools.py. A batch therefore needs an explicit file list and linked pull request so other contributors can identify its scope. The discussion's duplicate-work example makes the practical distinction clear: agreement on batching does not reserve a file already being changed in another pull request.
How should tool-refactor tasks be marked as claimed, in progress or completed?
joebog1 says ticked boxes mark claimed files in issue #3629, while another contributor uses checkmarks for completed or ongoing work. Write the status beside the filename and pull-request link to distinguish those meanings. The checkbox conventions in that discussion differ, so a checked item alone does not identify whether a migration is reserved, underway or merged.
Has the unified launcher proposal been approved for implementation?
Issue #2475 records an open launcher proposal and requests consolidation because several pull requests already exist. Its motivating problem names six scripts across four platforms with inconsistent names, flags and capabilities. The current setup guide gives platform-specific installation commands; the proposed launch and update interface belongs to the coordination discussion rather than the installation instructions used here.
What we have not tested yet
Our Docker attempt ended at the mount-safety check, leaving no completed chat task, research report, agent action, screenshot or known-answer result from that attempt. A corrected mount arrangement, GPU serving, email authorization, MCP execution, multi-user isolation, backup recovery and sustained load remain outside our completed work. The lab-node:22 label gives no actual runtime or operating-system version. The npm audit covers its dependency set, excluding a Python-dependency or application-security assessment.
Provider authentication has conflicting dated accounts. The August 21 independent review reports Gmail OAuth2 for Workspace and .edu accounts; the official setup guide checked on November 19 describes IMAP/SMTP username-password authentication and says Microsoft mailboxes generally require OAuth. Matching implementation evidence is absent, leaving current provider authentication unresolved. The independent review's sandbox assessment is a secondary reading rather than our security audit, and the ZIP discussion contains no malware analysis.
We have not reproduced the community workarounds or benchmarked the alternatives on a shared workload. Pearce describes starting Deep Research without a reproducible assessment of the finished report. The latest-release lookup returned no release data, which leaves release existence unresolved. Search demand and dependable daily deployment counts are unknown. A further evaluation needs one completed, checkable task through the exact model endpoint and integrations intended for daily use.


