SOLKRAFTMETHODS FOR AGENTS Not connected
OPEN SOURCE · BUILT FOR YOUR AGENT

Your words.
A working method.

Describe the work. SolKraft matches each requested capability to a reusable procedure, puts the steps in order, and gives your agent the instructions to follow.

Built for work that spans investigation, building, testing, and explaining the result. Ambiguous requests surface a need for clarification.

SoftwareResearchDataWritingBusinessMath + engineering
Local plugin · REST · MCP · CLI / Your agent keeps control
FROM INTENT TO INSTRUCTIONSREGRESSION EXAMPLE
YOUR REQUEST

“Inspect the repository, diagnose the CI failure, implement a fix, run regression tests, and write the PR description.”

↓ ONE REQUEST · FIVE METHODS
  1. 01
    Understand the codebaseTrace the source and its boundaries.
    ↗
  2. 02
    Diagnose the failureFind the cause before changing behavior.
    ↗
  3. 03
    Implement the fixApply the software build procedure.
    ↗
  4. 04
    Test the resultRun relevant regression checks.
    ↗
  5. 05
    Explain the changeWrite the pull request description.
    ↗
SOLKRAFT SELECTS → YOUR AGENT EXECUTES
173bundled procedures
473,000calls required by the local proof
7read-only MCP tools
Inspect the evidence ↗
THE ENGINE UNDER THE HOOD

Features

One request. A whole toolkit.

Intent becomes an inspectable workflow. Your agent gets the methods it needs, when it needs them.

Inspect→Build→Test
01 · INTENT → WORKFLOW

Five tasks. One sentence.

Compose ordered stages from a compound request: inspect → diagnose → build → test → explain. Cross domains without manually assembling a stack of skills.

⌘CODE◈DATA⌕RESEARCH✎WRITING∑MATH◎BUSINESS
02 · 173 PROCEDURES

A catalog with range.

Software, research, data, writing, business, mathematics, legal research, and engineering. Discover by capability and inspect the actual instructions.

CATALOG173 methods→SELECTEDFull instructions
03 · CONTEXT ON DEMAND

Load less. Get specific.

Search compact metadata first. Retrieve a selected skill and its supporting resources when needed, keeping the full catalog out of the conversation.

✓ Run tests⊘ Don’t deploy
04 · INTENT BOUNDARIES

“Don’t deploy” matters.

Separate active work from exclusions, quoted text, deferred tasks, and delivery context. Ambiguous stages surface a clarification need instead of a pile of plausible skills.

INPUTS→◇CONTRACT→OUTPUTS
05 · CONTRACTS BEFORE CONTENT

Know what a method requires.

Inspect inputs, outputs, effects, requirements, and trust metadata before loading instructions. Hardened routing rejects inadmissible capabilities. Your agent retains execution authority.

S✳
PLUGINMCPRESTCLI
06 · YOUR ENVIRONMENT

Local. Callable. Yours.

A native Codex plugin, seven read-only MCP tools, a CLI, and a REST API. Run locally without a hosted account or model API key for the service.

473,000 production-router callsCompleted local proof with downloadable receipts.100% clarified compound exact matchingZero extra selections and zero unresolved stages in the acceptance corpus.Original diagnostics preservedAmbiguous requests remain visible; the result does not claim universal language accuracy.
WHAT IT IS GOOD AT

Less hunting for skills.
More clarity about the next step.

Explore the catalog ↗

A practical layer between your request and your agent’s tools. Discover a method, inspect why it was selected, and retrieve its instructions when they are needed.

01 · ORDINARY LANGUAGE

Ask for the outcome.

Route actions and context without memorizing skill IDs. Scope exclusions, quoted text, deferred work, and delivery requirements are handled separately from active tasks.

02 · MULTI-STEP WORK

Keep the workflow visible.

Compose methods across domains, preserve requested order, and expose unresolved stages. Review the proposed route before your agent applies it.

03 · SELECTIVE RETRIEVAL

Load the method you need.

Search compact metadata first. Retrieve the selected procedure and supporting resources through REST or MCP, without putting the entire catalog in context.

Inspectable contracts Inputs, outputs, effects, requirements, and trust metadata.Hardened public routing Inadmissible capabilities are blocked and explained.Host-controlled execution A route supplies guidance; your agent’s tools and authority govern the work.
WHAT CHANGED · CONTRACT ARCHITECTURE

Skills are no longer opaque superpowers.

Read the architecture ↗

SolKraft now treats each skill as a compact, inspectable capability contract. The router can reason about what a skill needs, what it may affect, what authority a host would need, how skills connect through typed inputs and outputs, and what evidence would verify success—before loading the full procedure.

01 · CONTRACTS

Inputs, outputs, effects, verification.

Skills can carry typed contract.yaml sidecars with namespaced effects, capability/resource requirements, declarative checks, digests, provenance, and trust state.

02 · ROUTING

Admissibility before composition.

Known-inadmissible skills are filtered before they shape a route. Whole routes are validated, broken dependencies are exposed, and one bounded repair pass can substitute a safe alternative.

03 · DATAFLOW

Capabilities can compose by type.

Typed outputs can satisfy downstream inputs. A unique compatible producer can be inserted automatically; missing or ambiguous inputs remain visible instead of being guessed.

04 · TRUST

Verification is evidence-bound.

Execution evidence is bound to exact contract and SKILL.md digests. Trust comes from external digest-bound review, never from a skill declaring itself trusted.

05 · INTEROP

REST, MCP, CLI, OpenAPI.

Contract metadata is available over REST, MCP and CLI. Portable import/export and conservative MCP/OpenAPI adapters carry declarations only—never authority.

06 · CI

Pre-build a skill before the PR.

Contributors can run the same reusable full gate locally or from GitHub before opening a PR, including the 100,000-route regression, packaging, installed MCP checks, contract indexing, and hardening receipts.

Hardened public defaults Opaque skills fail closed over REST, MCP, and CLI. Lazy by design The route/index hot path still avoids loading full skill bodies. Execution stays external execution_authorized remains false.
✓ LOCAL EXACT-MATCH VALIDATION

The right methods.
Every requested step.

The full local gate requires 473,000 production-router calls across 32 slices: 173,000 single-skill cases, 100,000 clarified compound requests, 100,000 original diagnostic requests, and 100,000 stability calls. Results appear only when the complete exact-match receipt passes.

Download the receipt ↓
Loading completed receipt…NO BROWSER-SIDE CI
ROUTING CALLS—full proof + original diagnostics · 32 slices
SINGLE-SKILL DECISIONS—selection + policy behavior · 173,000 cases
COMPOUND TARGET COVERAGE—clarified requests · 100,000 cases
TARGET ORDER PRESERVED—expected methods in sequence
REPEAT STABILITY—10,000 base cases × 10 interleaved rounds
EXACT COMPOUND SKILL SET—all requested methods · zero extra selections
Exact matching is the acceptance gate.

Acceptance requires 100% exact compound sets, zero extra selections, and no unresolved stages in the clarified corpus. Original requests remain a separate diagnostic.

The receipt proves the stated corpus. Ambiguous original wording is reported separately; a generated test result does not establish perfect matching for arbitrary language, skill execution, or deployment.

INSPECT A REAL EXAMPLE

See the words.
See the methods.

Choose a recorded request routed by the same Python engine. The result is precomputed, inspectable, and available without a server.

Custom semantic requests require a connected SolKraft API. Browse and inspect bundled skills here without a connection.

production-router / recorded-exampleREAD-ONLY
Loading recorded request…
    No routing example loaded.
    What the validation does—and what it does not establish

    The generated corpus uses catalog-derived semantic anchors without injecting skill IDs or full descriptions. Its locked holdout is generated, not an independent human-language study. The authored regression suite checks everyday requests and scope boundaries separately. All calls use the production router with a frozen catalog; routing outputs are not memoized.

    Gates require 100% global, holdout, and per-skill decisions; 100% clarified compound coverage, ordering, and exact sets; zero extra selections and unresolved stages; 100% stability; zero metadata leakage; and complete matching receipts. Clarified requests retain every original topic and target, adding published capability words to distinguish ambiguous intent. All 100,000 unchanged original requests are rerun and disclosed separately.

    python -m scripts.local_ci --semantic-proof --semantic-workers 6
    Read the method and limitations ↗

    Source identity loads with the receipt.

    FROM A REQUEST TO USEFUL WORK

    Your agent does the work. SolKraft supplies the method.

    A skill is a reusable procedure with instructions and supporting resources. The solforge skill contains the orchestration and semantic routing machinery; it is also a node in the selection graph. Specialist skills supply the methods for research, building, testing, writing, and engineering. Capability-preserving execution supplies the adaptive execution loop: frame the outcome, map dependencies, schedule safe work, preserve evidence, repair failures, and verify completion. It is an instruction-based execution layer, not a hidden Python scheduler.

    FOLLOW THE REQUEST · A WALKTHROUGH, NOT A LIVE JOB
    01
    YOUR INTENT

    Describe the outcome.

    “Inspect the codebase, fix the bug, add a regression test, and verify the release build.”

    02
    SEMANTIC ROUTING

    Turn language into a route.

    SolForge reads actions, context, and exclusions. The selection graph supplies workflow relationships.

    03
    SELECTED PROCEDURES

    Choose methods that matter.

    Codebase investigation → software build → software test. The agent checks the route against your actual intent.

    04
    CONTEXT ON DEMAND

    Retrieve just the next skill.

    MCP or REST returns the selected instructions and needed references. The entire catalog stays out of the conversation.

    05
    CAPABILITY-PRESERVING EXECUTION

    Your agent does the work.

    The host preserves requirements, follows dependencies, applies the specialist procedure, and adapts when evidence changes.

    06
    VERIFICATION

    Show what actually holds.

    The agent runs relevant checks, repairs failures, and reports evidence and remaining limits. SolKraft does not certify the result.

    01

    Discover a procedure

    Browse the paginated catalog or search names and descriptions. Inspect a skill, then copy its instructions into your agent. This works here without an account or API key.

    02

    Understand the route

    Try the worked example below. For your own request, connect a SolKraft API and use Compose a route. The result shows stages, selected skills, reasons, and unresolved work.

    03

    Execute with preserved capability

    Copy instructions manually, or let an MCP or REST client retrieve them. Your agent reads the selected procedures, uses its own tools, and verifies the requested result.

    TRY IT · NO CONNECTION NEEDED

    A bug fix with a release check

    Inspect the codebase, fix the bug, add a regression test, and verify the release build. Do not deploy or publish anything.

    See the actual parser output generated from this request, inspect its selected procedures, and copy the request for your agent.

    How the semantic parser and graph actually work

    Identify the capability

    The deterministic parser still understands scoped rules, concept aliases, domain context, exclusions, quoted instructions, and completed or deferred work. It also matches published capability identity using distinctive tokens and unique capability signatures, matching active clauses while keeping excluded, quoted, deferred, and delivery-only text out of capability selection.

    Filter, compose, validate

    Hardened routing prefilters candidates through the ContractIndex before composition. The composer orders the requested stages, then whole-route validation checks contract admissibility, dependencies, and typed input/output dataflow. Bounded repair can recompose or repair a broken route instead of silently accepting an invalid chain.

    Keep the limits visible

    This is deterministic routing, not an LLM call. Scores are relative rankings, not probability estimates. Contract filtering, whole-route validation, and typed dataflow repair improve selection safety; a proposed route still never grants permission or executes a job.

    Catalog search on this static site is keyword filtering. Custom semantic composition runs in your connected server. The example is real, precomputed parser output and is labelled separately from a live response.

    Read the orchestration skill ↗ · Set up REST or MCP ↗
    PUT IT TO WORK

    Install once. Ask normally.

    Using Codex? Install the private plugin included with SolKraft. It lets Codex find useful skills, read their instructions, and use them while helping with your work. Install it directly from our GitHub repository using the steps below. You do not need Render, a hosted API, or a SolKraft API key.

    CODEX · LOCAL

    Install the included Codex plugin

    python -m pip install "git+https://github.com/halthinks/SolKraft.git"
    codex plugin marketplace add halthinks/SolKraft --sparse .agents/plugins --sparse plugins/solkraft
    codex plugin add solkraft@solkraft

    You need Python 3.11 or newer, Git, and Codex with plugin support. Run these commands in your terminal. The first installs the skill library; the next two add our GitHub repository to Codex and install its private plugin. Start Codex from the same terminal, then open a new chat. Do not search the public plugin directory for SolKraft.

    Installation and troubleshooting ↗
    OPTIONAL: HOST ONLINE

    Connect from another app or computer

    Host SolKraft on Render when you want another app or agent to connect to SolKraft over the internet. The included Codex plugin runs locally and does not need this.

    Direct ChatGPT connection has not been tested. To use a skill in ChatGPT, open it in Explore, copy its instructions, and paste them into your chat.

    Show me the hosting steps
    THEN JUST ASK

    “Fix this bug and test it.”

    The agent searches or routes the request, retrieves useful procedures, then edits and tests with its existing tools. For a release gate, pass that context so “draft PR” routes to software testing.

    This page is an explorer and setup console. It does not run your repository or claim to have fixed anything.

    Download the plugin ZIP ↓
    How it compares with other projects

    skills.sh focuses on discovering and installing skills. Superpowers supplies an opinionated development methodology and agent workflows. SolKraft combines a multi-domain catalog, deterministic compound-request routing, inspectable relationships, and selective retrieval through REST and MCP. These approaches can coexist; there is no measured claim here that SolKraft beats either project.

    FROM ZERO TO CONNECTED

    Pick your path.
    See every connection.

    Explore the steps, watch the signal move, and copy the settings you need. These are illustrated instructions; only the console’s Connect button makes a real API connection.

    ILLUSTRATED GUIDE
    WHAT SHOULD I USE?

    Plugin for Codex. MCP for agents. API for apps.

    The included plugin is the easiest way to use SolKraft in Codex. The API and MCP let other software use the same skill library. The plugin already uses MCP behind the scenes.

    USE IT IN CODEX

    Install the included plugin

    The plugin sets up SolKraft’s tools and tells Codex when they are useful. Codex can find skills, read their instructions, and apply them while working on your task.

    Install the Python package and plugin on your computer. You do not need Render or a SolKraft API key.

    Show me the plugin steps
    CONNECT AN AI AGENT

    Use MCP

    MCP is a standard way for an AI agent to discover and call tools. SolKraft’s tools let the agent search for skills, choose useful procedures, and retrieve their instructions.

    The local Codex plugin sets this connection up for you. Other agents need their own MCP setup. To connect to a SolKraft server over the internet, give your agent the server’s HTTPS /mcp/ address and private API key.

    Show me the agent connection steps
    BUILD YOUR OWN APP

    Use the API

    The API lets your website, app, or script ask SolKraft for skills through ordinary web requests. You can build a skill browser, retrieve instructions, or send a task to /v1/route to get suggested skills.

    This website uses the API when you connect a server and compose a route for your own request.

    Show me the API steps
    OPTIONAL: PUT IT ONLINE

    Host on Render

    Render hosts your SolKraft server so apps and agents on other computers can connect over the internet. That one service provides both the API and remote MCP connection.

    Choose this when another app or agent needs to connect over the internet. The local Codex plugin runs without it.

    Show me the hosting steps

    What happens when I ask for help?

    Say: “Find this bug, fix it, and test the result.” With the plugin installed, Codex can ask SolKraft for investigation, implementation, and testing skills through MCP. SolKraft returns the instructions. Codex reads them, edits your project, runs tests, and explains the result.

    SolKraft supplies the methods. Your agent does the work. You do not need to set up the plugin, API, and MCP separately. For local Codex use, install the included plugin and its required Python package.

    YOUR APIConnect to a SolKraft serverOptional · enables custom routing · your key stays in this tab
    THE WORKSPACE

    What are you working on?

    Connect to load catalog
    POPULAR SKILLSSELECT TO INSPECT
    ✳Your skill catalog is ready when you are.

    The bundled catalog loads automatically. An API connection is optional.

    ⌘
    BUILT FOR AGENTS

    Small context.
    Full capability.

    Search compact descriptions first. Pull the full instructions only when a skill earns its place in the workflow.

    Read the agent guide ↗
    ⟡
    ONE INTERFACE

    REST for apps.
    MCP for agents.

    One catalog serves a clean API and seven read-only MCP tools, including contract metadata and index inspection. Routing never grants execution authority.

    Connect your client ↗
    PROJECT STATUS COMMUNITY BUILT

    Early, open,
    ready to shape.

    Bring a skill, a bug report, or a better idea for how agents find their tools.

    Get support ↗
    solkraft://system-changelogBUILD SUMMARY · CURRENT
    $ solkraft --what-changed
    • Expanded the full local proof to 473,000 production routing calls, including the preserved original diagnostic corpus.
    • Fixed ordinary request routing, inline exclusions, inert quotations, deferred work, repeated test stages, and output requirements.
    • Preserved compound target order and enforced the requested skill limit.
    • Removed skipped ambiguous cases and duplicate composition templates from the proof harness.
    • Bound receipts to exact source inputs; incomplete, mixed-source, and smoke runs cannot pass the full proof.
    • Replaced simulated validation animation with actual receipts and recorded routing examples.
    • Made exact matching, zero extra selections, and resolved requested stages mandatory acceptance gates.
    • Kept contracts inspectable, public routing hardened, and execution authority with the host.
    [status] exact-match acceptance enforced · completed results require a verified receipt · execution remains external
    KEEP YOUR AGENT CURRENT

    One link. The latest SolKraft.

    Give your agent the update guide. It covers the runtime, bundled skills, plugin refresh, and connection checks.

    solkraft / agent-update.md ↗
    Read agent guide

    Use with the agent that manages your installation.

    Show the prompt

    Update my SolKraft installation using https://halthinks.github.io/SolKraft/agent-update.md . Inspect the installed source first, preserve local changes and configuration, update the runtime and bundled skills plus the native plugin where installed, then verify actual MCP tool calls. Report what updated and anything still requiring a restart.