# Rka Core > Research Knowledge Agent — MCP server + REST API + React dashboard for AI-assisted research orchestration - **Type:** MCP server - **Install:** `agentstack add mcp-rka-project-rka-core` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [rka-project](https://agentstack.voostack.com/s/rka-project) - **Installs:** 0 - **Category:** [Data & Analytics](https://agentstack.voostack.com/c/data-and-analytics) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [rka-project](https://github.com/rka-project) - **Source:** https://github.com/rka-project/rka-core ## Install ```sh agentstack add mcp-rka-project-rka-core ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About # RKA Core — Research Knowledge Agent [](https://github.com/rka-project/rka-core/actions/workflows/pytest.yml) [](LICENSE) **A local-first research operating system for turning day-to-day research activity into durable, auditable knowledge.** AI-assisted research produces valuable observations, experiments, decisions, and failures, but that reasoning is often scattered across conversations, repositories, notes, and terminal logs. RKA maintains a persistent, provenance-aware research record and helps researchers progressively transform it into claims, evidence clusters, research questions, and defensible publication structures. With RKA, researchers can: - resume long-running projects without reconstructing prior reasoning; - preserve why decisions were made and what evidence supports them; - separate tentative observations from reviewed claims; - supervise AI collaborators through explicit missions and decision gates; - expose auditable research records to downstream tools; and - support multiple research outputs from the same knowledge substrate without coupling those tools to Core. RKA is built for research workflows in computer science, AI, cybersecurity, IoT, and cyber-physical systems. The core is usable today; the interactive manuscript workbench and Agent-Native Research Artifact interoperability described below are active roadmap directions. ## Why RKA exists Research is not a sequence of isolated prompts. It is a long-running process of forming questions, trying approaches, collecting evidence, revising assumptions, and deciding what the evidence supports. Most AI tools preserve only fragments of that process. Conversation history remembers what was said but not necessarily what remains valid. Vector databases retrieve similar passages but do not explain why a decision was made. Task agents execute work but often discard the reasoning, alternatives, and failed branches that make results interpretable. RKA treats the research record itself as durable infrastructure. ```mermaid flowchart LR Inputs["Research activitynotes · conversations · files · repositories · experiments"] Record["Longitudinal recordobservations · decisions · failures · provenance"] Knowledge["Reviewed knowledgeclaims · evidence · clusters · research questions"] Spine["Publication reasoninginsight · gap · spine · contributions · evaluation"] Outputs["Research outputspaper · proposal · ARA · grounded dialogue"] Inputs --> Record --> Knowledge --> Spine --> Outputs style Record fill:#E6F1FB,stroke:#185FA5,color:#042C53 style Knowledge fill:#EEEDFE,stroke:#534AB7,color:#26215C style Spine fill:#FAEEDA,stroke:#854F0B,color:#412402 style Outputs fill:#E1F5EE,stroke:#0F6E56,color:#04342C ``` RKA Core owns the early stages of this pipeline: durable research records, retrieval, provenance, and reviewed knowledge. Manuscript development is an independent downstream capability in `rka-writer`; it is not activated by the Core distribution. ## How it works RKA coordinates three roles around one shared, typed knowledge base: - **Researcher** — frames the problem, contributes domain knowledge, ratifies consequential decisions, and retains final authority. - **Brain** — retrieves context, synthesizes evidence, maintains claims and research questions, proposes options, and keeps interpretations current. - **Executor** — performs bounded implementation or experimental missions, records findings, and raises checkpoints when assumptions or scope require review. - **RKA** — stores the shared record, provenance graph, lifecycle state, and auditable handoffs between them. The roles are architectural responsibilities, not requirements to use one particular model. RKA exposes its capabilities through MCP and REST. The reference workflows are currently tested most extensively with Claude Desktop and Claude Code; ChatGPT can connect through the authenticated HTTP MCP path. ### Progressive crystallization RKA does not require every raw note to become a formal conclusion immediately. Knowledge matures in stages: ```mermaid flowchart LR Journal["Journalobservations · procedures · directives"] Candidates["Interpretation candidatessource locator · uncertainty · falsifier"] Claims["Claimstyped assertions with source spans"] Scope["Claim scope versionsconditions · extension policy · falsifier"] Clusters["Evidence clustersrelated claims + synthesis"] Questions["Research mapquestions · gaps · contradictions"] Writing["Claim spineargument · contribution · evaluation"] Journal --> Candidates --> Claims --> Scope --> Clusters --> Questions --> Writing ``` Raw records remain available even when later interpretations change. Candidate interpretations must be reviewed explicitly before promotion; they can be deferred, rejected, merged, or classified without silently becoming scientific claims. Derived claims and clusters can be reviewed, superseded, or rebuilt without rewriting history. Each canonical claim has a separate, immutable applicability contract. Scope reviews record typed conditions, uncertainty, allowed and prohibited extensions, and falsifiers. Missing or stale scope stays visible rather than being inferred, and manuscript admission also continues to check grounding, scientific evidence status, contradictions, and freshness independently. ### Provenance by construction RKA connects literature, decisions, missions, findings, claims, and later decisions through typed relationships. A researcher can ask not only “What do we currently believe?” but also: - Which observations support this claim? - Which decision caused this experiment to be run? - What changed after an assumption was invalidated? - Which manuscript contribution still lacks sufficient evidence? - What did the project believe at an earlier point in time? ### Researcher-controlled agency AI collaborators may propose interpretations and execute approved work, but consequential commitments remain visible. Confirmation briefs, checkpoints, decision records, validation gates, and reversible supersession keep the researcher in control without forcing them to reconstruct every agent action. ## What RKA is—and is not RKA is not a replacement for a reference manager, a generic retrieval database, or a paper generator layered directly on top of chat history. It combines four concerns that are usually separated: | Concern | RKA's role | |---|---| | **Continuity** | Preserve research activity across sessions, tools, and collaborators. | | **Epistemic structure** | Distinguish observations, assumptions, hypotheses, results, and reviewed claims. | | **Control** | Record decisions, delegate bounded missions, and escalate uncertainty. | | **Publication grounding** | Build arguments from traceable claims and evidence rather than reconstructing provenance after writing. | Zotero, repositories, notebooks, and experimental platforms remain important sources. RKA links and interprets their outputs rather than attempting to replace them. ## What is available today - **Persistent, multi-project research records** for journals, literature, decisions, missions, reports, checkpoints, and artifacts. - **Interpretation staging** with exact source locators, uncertainty, falsifiers, immutable review history, and explicit promotion or revocation. - **Canonical claim-scope contracts** with immutable revisions, typed applicability conditions, extension policy, falsifiers, and fail-closed manuscript readiness. - **Claims and evidence clusters** with source-span provenance, confidence states, contradictions, and review workflows. - **Research maps** connecting research questions to clusters and individual claims. - **Decision and freshness lifecycles** with supersession, staleness propagation, assumption tracking, and historical belief queries. - **Brain–Executor workflows** with scoped missions, acceptance criteria, backbriefs, checkpoints, and reports. - **Hybrid retrieval and graph navigation** using FTS5, optional vector embeddings, typed links, and multi-hop context assembly. - **Researcher-facing dashboard** for browsing projects, journals, decisions, missions, research maps, provenance, and audit history. - **Role-specific skills** for strategic research management, execution, and PI supervision. - **MCP, REST, CLI, and knowledge-pack interfaces** for local and connected workflows. Detailed feature and workflow documentation lives in the [User Manual](docs/USER_MANUAL.md) and [Usage Guide](USAGE_GUIDE.md). ## From one research record to papers and ARA RKA's publication direction is based on a simple principle: researchers should not have to reconstruct their scientific reasoning separately for every output. The separately developed manuscript workbench will guide a researcher from an initial insight through problem scoping, related-work positioning, gap analysis, challenges, innovations, research questions, contributions, evaluation design, claim spine, outline, and full draft. Throughout that process, proposed text remains linked to RKA claims, evidence, decisions, and source records through Core's public contract. This creates a natural interoperability point with the [Agent-Native Research Artifact](https://github.com/ARA-Labs/Agent-Native-Research-Artifact) project: - **RKA** provides the longitudinal capture, reasoning, review, and authoring environment. - **ARA** provides a portable agent-native research package spanning scientific logic, executable assets, exploration history, and evidence. - **Traditional papers and ARA packages** can become two grounded views of the same research effort instead of two separately maintained records. The intended integration is an explicit, testable crosswalk—not a lossy text export. Stable RKA entities should map to ARA objects with preserved provenance, lifecycle status, and claim–evidence bindings. This interoperability is planned work and is not yet part of the released core. ## Quick start ### 1. Install and start RKA Prerequisites: Git, Python 3, [uv](https://docs.astral.sh/uv/getting-started/installation/), and Docker with the Compose v2 plugin (`docker compose version` must work). Docker Desktop includes Compose on macOS, Windows, and Linux; Linux users may instead install Docker Engine plus the Compose plugin. macOS or Linux: ```bash git clone https://github.com/rka-project/rka-core.git cd rka-core docker compose up -d uv tool install --force --reinstall . ~/.local/bin/rka --version curl http://127.0.0.1:9712/api/health ``` Windows PowerShell: ```powershell New-Item -ItemType Directory -Force "$env:USERPROFILE\Code" | Out-Null Set-Location "$env:USERPROFILE\Code" git clone https://github.com/rka-project/rka-core.git Set-Location rka-core docker compose up -d uv tool install --force --reinstall . & "$env:USERPROFILE\.local\bin\rka.exe" --version Invoke-RestMethod http://127.0.0.1:9712/api/health ``` Open [http://127.0.0.1:9712](http://127.0.0.1:9712). The interactive REST documentation is at [http://127.0.0.1:9712/docs](http://127.0.0.1:9712/docs). The MCP binary is `~/.local/bin/rka` on macOS/Linux and `%USERPROFILE%\.local\bin\rka.exe` on Windows. Client-specific Claude and Codex configuration, upgrades, and troubleshooting are documented in [INSTALL.md](INSTALL.md). > The first uncached FastEmbed startup downloads the embedding model. During that download or a generation rebuild after an upgrade/import, health remains available but semantic search can temporarily fall back to lexical retrieval. Check Settings for indexing progress before judging retrieval quality. ### 2. Begin a project After connecting a client, ask it to list or create an RKA project and state the selected project at the beginning of the session. Every project-scoped operation uses an explicit project ID so work cannot silently land in the wrong project. For a complete first-project walkthrough, see [USAGE_GUIDE.md](USAGE_GUIDE.md). ## Interfaces | Interface | Best for | Documentation | |---|---|---| | **Web dashboard** | Browsing, reviewing, navigating, and direct editing | [User Manual](docs/USER_MANUAL.md) | | **MCP** | AI-assisted research retrieval, maintenance, and execution workflows | [Installation](INSTALL.md), [Technical Reference](docs/TECHNICAL_REFERENCE.md) | | **CLI** | Starting services, status, backup, credentials, and workspace bootstrap | [Technical Reference](docs/TECHNICAL_REFERENCE.md) | | **REST API** | Custom integrations and application development | [Technical Reference](docs/TECHNICAL_REFERENCE.md), live `/docs` | | **ChatGPT connector** | Authenticated access from ChatGPT to a local RKA instance | [Connector Guide](docs/CHATGPT_CONNECTOR.md) | | **Writer (separate project)** | Explicitly invoked manuscript assistance using RKA's public contract | [`rka-project/rka-writer`](https://github.com/rka-project/rka-writer) | ## Architecture The core distribution runs a FastAPI service and React dashboard, a background indexing worker, and an MCP adapter over a shared service layer. SQLite, FTS5, and sqlite-vec provide local persistence and retrieval. Server-side research interpretation is intentionally separated from storage: the connected Brain performs synthesis while RKA preserves, validates, and serves the structured record. See [Architecture](docs/ARCHITECTURE.md) for the runtime model, data layers, provenance graph, MCP dispatch surface, search design, and extension boundaries. ## Roadmap The immediate priority is to harden RKA as an independently usable research- knowledge core. Writer remains a separate downstream product under the same RKA Ecosystem project: 1. **Core reliability** — harden journal correctness, provenance, project isolation, retrieval, claim edges, migrations, export/import, and recovery. 2. **Stable integration contract** — freeze the supported REST/MCP surface and provide version, capability, and compatibility discovery. 3. **RKA Writer extraction** — move the Writer skill, manuscript semantics, academic-writing tools, and Workbench to `rka-project/rka-writer`. 4. **Future Core slimming and ecosystem validation** — slim the Core distribution only in an explicitly approved breaking release, after legacy-state migration and downstream compatibility tests pass. The earlier Agentic repository/extraction proposal is shelved. Its history is preserved, but it is not an active product, dependency, or installation path. The Workbench, ARA interoperability, and dual-output evaluation remain planned Writer/ecosystem work; they are not the immediate Core implementation target. The dependency-ordered plan lives in the repository [Roadmap](ROADMAP.md), with active work tracked through [GitHub milestones](https://github.com/rka-project/rka-core/milestones). Roadmap items describe direction and should not be interpreted as released features. ## Research The working paper [*Framing Is Human: Researcher–Brain–Executor Architecture for AI-Assisted Research*](docs/paper/RKA-paper.pdf) describes RKA's architectural argument, design principles, and evaluation. RKA is being developed for research workflows at UNC Charlotte. Feedback, comparative evaluations, interoperability experiments, and research collaborations are welcome. ## Documentation | Document | Purpose | |---|---| | [Installation](INSTALL.md) | Complete local, MCP-client, and connector setup | | [Usage Guide](USAGE_GUIDE.md) | End-to-end Brain, Executor, PI, and research workflows | | [User Manual](docs/USER_MANUAL.md) | Concepts, dashboard operation, and researcher-facing reference | | [Architecture](docs/ARCHITECTURE.md) | Design rationale, components, data model, and knowledge lifecycle | | [Technical Reference](docs/TECHNICAL_REFERENCE.md) | CLI, MCP, REST, configuration, and development entry points | | [Core Profile](docs/CORE_PROFILE.md) | Supported Core dependencies, test boundary, and startup smoke gate | | [Roadmap](ROADMAP.md) | Dependen … ## Source & license This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [rka-project](https://github.com/rka-project) - **Source:** [rka-project/rka-core](https://github.com/rka-project/rka-core) - **License:** MIT Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** yes - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/mcp-rka-project-rka-core - Seller: https://agentstack.voostack.com/s/rka-project - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.