AGENT-NATIVE WORKSPACE
Built for AI agents, not just for humans.
Give your agent the context behind the ticket
Connect the agent you already use to Specky's Product Graph. OAuth-capable MCP clients need no pasted key; every action still lands in a workspace-scoped, human-reviewable flow.
- MCP and REST interfaces
- Context-aware product work
- Human review stays in control
The ways your agent talks to Specky
Use MCP for tool-using agents, REST for scripts and services, agents.json for discovery, or webhooks when you want Specky to notify your system.
From zero to first action in 3 minutes
Choose OAuth or a workspace key
For Claude Desktop, ChatGPT, and other OAuth-capable clients, paste the MCP endpoint and approve the browser prompt. For a headless CLI or automation, create a workspace-scoped sk_live_… key in Settings → Developer. The feedback_to_spec tool also works without either for a quick test.
Point your agent at Specky
Add https://www.specky.space/api/mcp to your client, or use /openapi.json for REST. OAuth clients discover authorization automatically; key-based clients send Authorization: Bearer sk_live_….
Act on the user's behalf
Read get_agent_context and get_engineering_standards first, then search the graph, draft a PRD, generate tickets, or kick off research. When a ticket is handed to an agent, validate the plan before coding, validate changed files before success, and use list_agent_runs/update_agent_run to keep branch, commit, PR, CI, deployment, blockers, and outcome visible.
curl -X POST https://www.specky.space/api/mcp \
-H "Content-Type: application/json" \
-d '{
"jsonrpc": "2.0",
"id": 1,
"method": "tools/call",
"params": {
"name": "feedback_to_spec",
"arguments": { "feedback": "paste a real piece of customer feedback here" }
}
}'curl -X POST https://www.specky.space/api/extension/capture \
-H "Authorization: Bearer sk_live_…" \
-H "Content-Type: application/json" \
-d '{
"content": "Competitor launched a new pricing page with usage-based tiers.",
"url": "https://example.com/pricing",
"source_type": "COMPETITIVE_INTEL"
}'Want to see what to build before you configure every tool? Pick a blueprint for feedback, research, retention, reporting, or an n8n automation — each one shows the input, the handoff, the human approval point, and the output.
Bring your own agent. Claude, Codex, or whatever you use next.
One endpoint and the same product tools in every harness. OAuth-capable clients open a browser so you can choose the workspace; headless scripts can use a workspace key.
https://www.specky.space/api/mcpStreamable HTTP (MCP spec 2025-06-18)
· Auth: OAuth 2.1, or Bearer sk_live_… for headless clients
curl -X POST https://www.specky.space/api/mcp \
-H "Content-Type: application/json" \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'Claude Code
Anthropic’s coding agent in the terminal, IDE, and web
claude mcp add --transport http specky \ https://www.specky.space/api/mcp \ --header "Authorization: Bearer sk_live_your_key"
Then ask it to search your Product Graph, read an approved PRD, and implement against the cited evidence.
Codex
OpenAI’s coding agent (CLI, IDE extension, and cloud)
# ~/.codex/config.toml
[mcp_servers.specky]
url = "https://www.specky.space/api/mcp"
http_headers = { Authorization = "Bearer sk_live_your_key" }Codex reads the same tools Claude Code does. Running both against one approved spec is a cheap way to find out whether the spec is underspecified.
Claude Desktop & the Claude apps
Anthropic’s chat clients, via custom connectors
Settings → Connectors → Add custom connector URL: https://www.specky.space/api/mcp No key. Approve the browser prompt, pick a workspace, done.
Specky is an OAuth 2.1 authorization server, so the Claude apps run the browser consent flow themselves — nothing to paste. Revoke any time in Settings → Developer → Connected apps.
Cursor
The AI-native editor
// .cursor/mcp.json
{
"mcpServers": {
"specky": {
"url": "https://www.specky.space/api/mcp",
"headers": {
"Authorization": "Bearer sk_live_your_key"
}
}
}
}Query the graph mid-edit — “what did customers actually say about this export flow?” — without leaving the file.
ChatGPT
Custom connectors in developer mode
Settings → Connectors → Add custom connector URL: https://www.specky.space/api/mcp Auth: OAuth (discovered automatically)
No key to paste — Specky advertises its authorization server at /.well-known/oauth-authorization-server and ChatGPT runs the consent flow. Whether custom connectors are available at all depends on your ChatGPT plan and workspace settings.
n8n
The self-hostable automation runner
MCP Client Tool node Endpoint: https://www.specky.space/api/mcp Transport: HTTP Streamable Auth header: Authorization: Bearer sk_live_your_key
Lets an n8n workflow call Specky tools directly instead of hand-rolling REST calls — and lets an n8n AI agent use the graph as memory.
Also works with: Windsurf · Zed · VS Code agent mode · Cline · Goose · Continue · LibreChat · Any client speaking MCP Streamable HTTP. “Verified” means we’ve run it against the server ourselves; “compatible” means it speaks the transport and should work — tell us if it doesn’t.
What your agent can do
These are the tools available to agents today. MCP exposes the richer tool surface; REST covers the core workspace actions.
feedback_to_specNo API key required. Turn one piece of raw feedback into an evidence-backed insight + PRD problem statement — call it straight from tools/list with no Authorization header at all.
search_workspaceHybrid ILIKE + pgvector search across every signal.
read_workspace_itemFetch any graph node by id — doc, ticket, insight, decision, OKR.
get_agent_contextRead the central build contract and product context before planning or coding. Secrets and unrelated preferences are excluded.
get_engineering_standards / list_engineering_rulesRead the shared definition of done and relevant MUST/SHOULD/MAY engineering rules for the repository and stack.
validate_agent_plan / validate_agent_changeRun transparent engineering gates before coding and before reporting success; findings are saved to the execution ledger for human review.
list_agent_runs / get_agent_runFind the provider-neutral execution record for a ticket handoff, including the current next action, CI state, deployment state, and attached artifacts.
update_agent_runReport progress, branch, commit SHA, pull request, checks, deployment environment/URL, blockers, and completion from any coding agent.
GitHub delivery syncPR, GitHub Actions workflow/check, and deployment_status webhooks update the same run and mirror delivery state back to the PM ticket. A merged PR is not treated as production proof.
create_documentDraft PRDs, research docs, and notes. Auto-indexed into the graph.
create_ticketsGenerate engineering tickets with acceptance criteria. Push to Jira, Linear, GitHub, Trello, or Shortcut.
create_insightRecord customer insights with evidence links.
create_opportunityAdd to the Opportunity Tree under a chosen outcome.
create_experimentSpin up A/B tests, fake doors, prototype tests.
create_validation_pageTurn an experiment into a typed public page that captures responses and feeds evidence back into the graph.
get_validation_page_analyticsRead variant funnels, confidence intervals, sample-size readiness, privacy-safe segments, guardrails, and a cautious experiment verdict.
publish_validation_page_to_repositoryCreate or update a specky/validation/... GitHub branch with a working fake-door artifact, manifest, README, and optional draft PR.
create_decisionLog strategic decisions with rationale and citations.
create_okrCreate OKRs with key results tied to the graph.
start_research_sessionLaunch an Alex JTBD campaign with a shareable link.
run_custom_workflowExecute a workspace-defined multi-step workflow.
list_inboxSurface items awaiting human review.
How to be a good agent in Specky
Human-in-the-loop by default
Create items in draft / proposed state. Specky's PM Inbox is the review queue — auto-publish only when the user explicitly asks for it in the same turn.
Cite the graph
When drafting PRDs, insights, or decisions, link evidence_node_ids back to source nodes. Specky surfaces citations inline so PMs can audit your reasoning.
One credential, one workspace
Every OAuth grant and API key is scoped to one workspace. Treat that product data as confidential, do not send it to services the user did not choose, and back off when Specky returns 429.
Read the build contract first
The get_agent_context and get_engineering_standards tools return the workspace's implementation rules, product context, principles, source context, persistent facts, definition of done, and escalation rules before an agent starts work. Custom rules can be selected by collection, tag, language, and framework.
Give your agent a useful first job.
Start with one real product question, then let the agent search, draft, and stage the next move with its evidence attached.