YOUR JUNCTION WORKBENCH
All the tools. One room.
Connect a perk, try an idea, build something real.
Service health
Checks saved credentials and makes a small demo request for each connected LLM service. Uses provider credits. Failed services are disabled until a successful recheck; temporary rate limits are reported separately. External workflows require manual verification.
Resources in your kit
Tools ready
Requests today / 500
No tools match your search.
A home for your next idea.
Create a project to brainstorm with your assistant, keep Markdown notes and map the pieces of your plan.
Main project LLM
Applies to chat, project updates, task suggestions, plans, reviews and implementation. Shared history and selected context go to the saved provider and use its connected account credits. Verda requires an inference API key. Choose a text model with enough context; availability does not guarantee compatibility. Manual selections stay fixed. Auto chooses an eligible provider/model per task using Model checker evidence.
From suggestion to review.
Your assistant plans with your selected code, tools and notes. Upload updated code before asking it to review your implementation.
Implementation changes
Explicit implementation requests can edit imported copies when the project implementation flag is enabled. Original local files and remote Git repositories remain unchanged. Export the diff to apply it yourself.
Source references
Import code for assistant to read. These are project copies, not live folders. Selected files are sent to the selected provider with chat, planning and review requests.
Server path
Read-only references. Refresh to pull disk changes into project copies. Conflicting edits are preserved. Server originals are never changed.
Local files
Up to 30 files per upload, 100,000 characters each. Dependency folders, .git, .env and key files are excluded. A folder is copied once; upload again after changing it.
GitHub repository
GitHub HTTPS repositories are supported. Imports are pinned to the browsed tree. Other Git hosts can be added using local file upload. Re-browse to pull newer code.
Imported files
View / edit a project source copy
Tools for this project
Junction simulator context
When enabled, each chat or task-model request reads the latest saved state of this run. Project swarms snapshot it once at start. Includes run statistics, challenge briefs, up to 20 teams and recent events; team/player filters add focused observations. This is read-only reference data, not permission to control the simulation. A failed refresh blocks the answer instead of silently using stale data.
Preview the simulator data sent to the model
Context disabled.
Select the services this project should use. The assistant sees the selection when planning. You can run supported operations here and keep their results as project context.
Off by default. Enabling this exposes an explicit Implement action for each task. It can edit up to three selected files and records a diff. It does not run code, tests, Git commits or pushes.
Run a project tool
Uses connected credentials and provider credits. Calls run only when you click Run; the latest five saved results are included in assistant’s project context.
No result yet.
Build your project with AI.
gpt-6-astraThe assistant knows all tool manuals, setup status, service notes, comments and perk dates. Sending a question saves your edits and shares this workspace’s notes, mind map, selected source files, task list, tool results and recent chat with the selected provider. Chat history is saved here for signed-in operators. Uses your selected provider’s credits; keep secrets in connection fields.
Markdown notes · up to 30,000 characters
Preview
Connect ideas into a plan. Select a node to edit it.
Up to 100 ideas / 10 levels. Save project to keep changes. Assistant suggestions can be added as nodes from chat.
Every chat includes all service manuals, capabilities, setup status, saved notes, comments and perk dates. Credentials stay on the server. Chats are saved and shared with signed-in operators. New chats automatically check connected providers and select a responding text model (uses provider credits). Each conversation has its own notes, sources, tasks, tool results and model selection.
Create a task swarm
Specialists share a private task board, use connected models and produce a final answer. Members take turns; the first member coordinates the final synthesis. Runs and contributions are saved.
Private task board
Ask the best-fit available model
Uses connected providers and exact model matches to published benchmark evidence. Automatic decisions use Gemini first, Featherless second, and OpenRouter/Jev last; the local option uses keyword rules without an inference call. Rankings are estimates, not guarantees. The selected model answers your task using its connected service. Answers, connection checks and model decisions use credits.
Answer
Recommendation
Automatic selection sends each task to the selected provider, with Jev through OpenRouter for classification when available. It uses the balanced/quality/cost/fast policy you analyzed. This changes the workspace LLM only; implementation permissions remain unchanged.
OpenAI-compatible API
Base URL: https://tools.188.166.112.226.nip.io/v1. Use your unified Toolroom bearer key. Select toolroom/auto or an exact provider/model ID from GET /v1/models.
POST /v1/chat/completions
{"model":"toolroom/auto","messages":[{"role":"user","content":"Review this Python function…"}],"router":{"strategy":"balanced","engine":"auto"},"max_tokens":512}Text-only Chat Completions subset, one answer, up to 2,048 output tokens. SSE is buffered after generation. Images, function calls, tools, Responses endpoints and unsupported parameters return an explicit error. API requests use only submitted messages, not saved project data. Routing details appear in the response's toolroom field.
Terms recorded in the Junction brief, 13–14 November 2026. Account expiry and balances have not been verified. All times below are UTC.
Perks with a countdown
Need an account date
Non-expiring grants
Upcoming dates
Only dated entries appear here. Event and writing deadlines are separate from perk expiry. Bars show time remaining from now; dates outside the view retain their full countdown.
Partner perks & availability
A Closed listing or a promotion quarter does not establish a credit expiry. Add the actual redemption/start date or exact expiry from your account to start a countdown. Calculated dates are estimates until confirmed with the provider; credits can run out sooner.
The last 100 requests. Payloads, results and secrets are not retained.
| Time | Tool / operation | Result | Latency |
|---|
Consilium · case memory and agent advice
Classify feedback such as possible disappointment, corrections or blockers. A specialist reviews each case; repeated incidents can trigger a second reviewer who reads and challenges the first review. Advice is supplied at the agent’s next model call, not injected into an already-running provider request.
Optional and off by default. Stores redacted message excerpts and preceding assistant context, shared with workspace operators. Enabled classification and reviews use connected providers and credits. Classification is an inference, not a fact about the user. Case memory is advisory and does not grant implementation or tool permissions.
Report an incident to a running agent’s context
Use the scope shown on an existing event, project:PROJECT_ID for a project/helper, or swarm:SWARM_ID. Codex scopes are shown below; API clients can supply metadata.toolroom_scope. Each scope has separate case memory.
Cases and lessons
Resolve a case to stop supplying its advice. Dismiss a false positive to suppress further advice for that case. Edit a lesson to preserve your correction during future reviews. Delete removes the case and its linked incident excerpts.
Classified events
BUILD ON YOUR TOOLKIT
One API key. All connected tools.
Use a dedicated bearer token from your app or scripts. Calls use the credentials you connect here and share this workbench’s request limits.
Checking service key…
One service key authenticates every API tool; provider keys stay on the server. Console/CLI tools remain external workflows. A token grants full workbench access. Replacing it immediately revokes the previous token.
Example: search with Exa
Discover operations with GET /api/tools. Send POST /api/run with a tool ID, an operation and a params object. The tool panel provides editable examples for every operation.
Limits: adaptive concurrency and pacing per service, provider request/token quota windows, shared daily allowance, 60-second upstream timeout (90 seconds for the isolated Codex worker) and 2 MB responses. API model chat outputs are capped at 2,048 tokens; Codex treats the requested output size as guidance, not a hard token ceiling. Request limits are not dollar budgets; provider billing still applies. No automatic retries.
Codex request logs
Optional, off by default. Includes prompts, code, tool arguments/results and model replies sent through /v1/responses. Authentication headers are never recorded; known saved credentials are redacted. Other sensitive content may remain in prompts and files.
Completed UTC hours are compressed automatically into ZIP files. The current hour stays as JSONL until the next hour. Archives stay on the server until manually removed. Only signed-in operators can access them.