DevTrellis MCP server
Commit Nov 12, 2026Allows AI coding agents to lookup, search, and read source-quoted technology research to reduce documentation re-reading and token usage.
Also known as nihatbirbudak/devtrellis
- Stars
- 0
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Nov 12, 2026
- License
- MIT
What is DevTrellis MCP?
DevTrellis allows AI coding agents to lookup, search, and read source-quoted technology research to reduce documentation re-reading and token usage.
Summary generated from the project's README
About DevTrellis MCP server
Developers and AI coding agents can reduce documentation re-reading and token consumption by accessing a shared library of technology research. This MCP server allows an assistant to check if a specific documentation page is already covered using the lookup_source tool before attempting to fetch it. By searching and reading existing records, the agent can find structured data and source-quoted research that has already been processed.
The system operates locally on your computer, keeping private research and metadata secure. While reading stored content does not require a model account, you can connect optional providers like NVIDIA, Claude, or Codex to perform summaries and deeper analysis using your own API keys or signed-in CLIs. The server is built with Python 3.12 and supports English and Turkish interfaces. Users can choose to keep research local or send anonymous, moderated copies to the community service to help build a collective knowledge base.
What you can do with it
- 1Prepare research contributionsUse the Claude Code or Codex skills to help an AI agent draft a new research report or plan a deeper investigation.
Written from the project's README. Every use case is checked against a quote from it.
README
View on GitHubDevTrellis
Discover. Research. Build together.
Türkçe · MIT license · Research & free NVIDIA setup · Community service
DevTrellis is an open technology knowledge community for hobbyists and professional developers. Its aim is to make useful research easier to find and reuse, helping reduce the time and tokens spent asking AI to investigate the same topic again. By bringing together current news, research, projects and developments, the community can also help people discover work and topics they did not know to look for.
Start at devtrellis.org: open a title or choose Read research. A shared report opens as a readable page; a record published through the machine gate opens as structured data (JSON) with its quotes and sources, because these records have no readable page yet. Follow the original source links. Check its date, sources, scope and limitations. If it answers your question, reuse it; if something is missing or outdated, research that part with your own model. Reading on the site needs no account, installation, API key or AI assistant. Reading and importing an existing report do not call a model.
Community pages: About · Guidelines · Volunteer application · Support. All offer English and Turkish without an account. Applications and support messages are private; applying is not required to read or share research.
Your personal research stays on your computer. When it is ready, choose Keep local or Preview and share: review the exact public text, sources, model details and content permission before sending an anonymous copy for moderation. On devtrellis.org, records that pass the machine gate are published automatically and labelled "not human-reviewed"; reports shared from the local app go to a moderator. Private question/claim metadata and author identity are excluded. Sharing is optional and requires neither an account nor manual file transfer. Finding, checking, reusing and contributing should build a useful shared library. The product vision and roadmap distinguishes current features from planned work.
Two ways to use it
Research can develop through linked contributions: cite the exact earlier report version, explain new findings and keep its original sources. Readers can follow that history and report usefulness, actual reuse, outdated information or a disputed claim. Separate web and MCP access counters show approximate activity for shared reports (machine-gated records are not counted yet); popularity and model names do not establish correctness. Research development and privacy.
This is an early alpha, previously named UFUK. Legacy UFUK_* environment variables, database keys and the ufuk-share format remain compatible. The community service is installed and operated separately. Living topic pages, an hourly community researcher and publisher connectors remain planned.
| Personal workspace | Community service |
|---|---|
| Follow sources and save interests, projects and research locally | Browse published research, projects and developments |
| Use your own optional model provider | Read without a model account or contribution requirement |
| Preview and send selected research directly; keep everything else local | Receive account-free contributions; machine-gated records can be published automatically (labelled "not human-reviewed"), shared reports go to moderators |
| Keep private research on your computer | Version corrections and withdraw publications |
The developer space, advanced workspace and community screens offer English and Turkish. Source and contributor text keeps its original language; interface language does not rewrite other people's work. Model report language is a separate provider setting.
An optional MCP connection lets an assistant check whether a documentation page is already covered (lookup_source) before fetching it, search and read records, and offer a contribution with an explicit preview and approval. Record text returned to an assistant still consumes context tokens. On covered questions, early experiments suggest that answers of similar quality can be reached with fewer page opens. Limits: 6 questions (11 November 2026, alpha.14 bridge), chosen to match existing records; 2 runs per question; the bridge was called by a script, not from inside a real assistant; one evaluator; an artificial cold baseline that read whole pages as raw files; billed cost and general savings were not measured; against WebFetch no saving was shown.
An assistant needs only the MCP connection to read records; the optional Claude Code / Codex skill helps prepare a contribution or plan deeper research. Compatible MCP Apps hosts show Share / Keep local controls; other hosts provide the same controls on a local review page. Submission requires the review capability and two initially unchecked confirmations. A model-supplied approval boolean is insufficient. Pattern filters reduce accidental disclosure, but cannot detect every private fact; inspect the exact public copy. Skill installation and privacy boundaries.
Install
Python 3.12 is the tested runtime. Node.js is only needed for JavaScript tests. No Docker, administrator account or model subscription is needed for local browsing and drafts. Run from the extracted source directory on Windows:
python -X utf8 kur.py
.\.venv\Scripts\python.exe -X utf8 calistir.py sun --elle-toplama
Open http://127.0.0.1:8781/platform. Stop with Ctrl+C. Setup creates .venv, installs requirements-lock.txt and checks dependencies without deleting application data. python -X utf8 kur.py --check checks an existing environment without installing anything.
Unix equivalents are python3 kur.py and .venv/bin/python calistir.py sun --elle-toplama. See release notes for tested operating systems; a command example is not platform validation.
--elle-toplama disables scheduled collection and research. A fresh installation has no community contributions. To fetch configured public sources explicitly:
.\.venv\Scripts\python.exe -X utf8 calistir.py topla
Collection makes outbound requests. Reading stored content does not spend model quota. Starting without --elle-toplama enables scheduled collection and the configured research workflow; review settings first. The local service binds to 127.0.0.1 for one person. Shared access uses the separate community service. Windows desktop launchers retain the legacy names RADAR-BASLAT.cmd and RADAR-DURDUR.cmd.
Data and updates
The default personal database is data/radar.sqlite3. To keep data outside extracted release folders, use a stable --db path for every command:
.\.venv\Scripts\python.exe -X utf8 calistir.py yedekle --db C:\DevTrellisData\personal.sqlite3
.\.venv\Scripts\python.exe -X utf8 calistir.py sun --elle-toplama --db C:\DevTrellisData\personal.sqlite3
Before updating, stop your instance and take a consistent backup with yedekle. Extract the new release separately, install its dependencies and use the same database path. Keep the old release and pre-update backup. Rollback after a schema change uses the pre-update backup; preserve the newer database separately so later edits are not discarded.
Do not publish data/, backups, .deps/, .venv/, logs, keys or private research. Release archives contain source and tests, without personal data or the parent repository's history.
Optional model providers
- NVIDIA: summaries, screening and source-based research with your own API key. Free prototype access is available under NVIDIA's account/model conditions; see the step-by-step guide. Use
nvidia_anahtar.pyfor hidden key entry, environment variablesUFUK_NVIDIA_API_KEY/NVIDIA_API_KEY, or Windows Credential Managerufuk-nvidia/api-key. - Claude: research using your own signed-in CLI;
UFUK_CLAUDE_CLIselects an executable when needed. - Codex: explicitly selected analysis of sources collected by DevTrellis, using your own signed-in CLI. This does not promise independent web research. The adapter restricts tools and supplies only the selected context. Install/sign in separately following the official authentication guide. If detection fails, set
UFUK_CODEX_CLIto the native CLI executable; Windows requires a.exe, not a command shim. Model selection can use your existing CLI setting.
Access, limits and charges are your responsibility. Requests send selected context to that provider; credentials are not included in community sharing. Accounts are not pooled community quotas. Without a provider, reading, source management, drafts, backups and card exchange remain usable. Release notes distinguish simulated adapter tests from live model calls.
Sharing and licenses
Save a card, choose its content permission and preview all fields before exporting. Research needs a source link; projects need a public project link. Importing previews before saving. Uploading to a community installation is a separate explicit action.
The optional K1 v4 contribution path (machine gate, moderation queue, stop switch) is disabled by default in the code; devtrellis.org runs it with automatic publication on, so records that pass the gate go live labelled "not human-reviewed". See docs/community-k1.md.
Software is MIT licensed, including commercial use. Authors independently choose CC BY 4.0 or CC0 1.0 for their own card text. These permissions do not change linked software licenses or source article rights. Link to publications and write your own summary. Account identity, author claims, model output and factual verification are distinct.
Development
.\.venv\Scripts\python.exe -X utf8 -m unittest discover -s tests -p 'test_*.py'
node --test tests/*.cjs
Tests use temporary data and fake providers unless documented otherwise; live probes are excluded from release archives. GitHub Actions runs the standard Python and JavaScript suites on Ubuntu and Windows with Python 3.12 and Node.js 22. Optional real Windows launcher tests remain a separate local acceptance step. See CONTRIBUTING.md, SECURITY.md, PLATFORM.md, and THIRD-PARTY-NOTICES.md.
The optional on-demand source memory pilot is disabled by default and uses a separate database. It introduces no periodic crawler or automatic model calls.
DevTrellis MCP questions
Is my research data kept private?
Yes, your personal research stays on your computer and you can choose to keep it local or preview and share it anonymously.