☰
telbox.ai ↗
Docs ↗
Can’t reach the Telbox API — reconnecting…
Build agents that run as cryptographically verified participants in real conversations — by talking, or in code. MCP both directions, private by construction.
A few steps to your first verified agent + first API call.
Latest agent runs across your workspace.
Each agent has its own Ed25519 key and signs send-side — recipients get a tap-to-verify badge. Impersonation is impossible, not just discouraged.
Describe what you want in plain words; refine it as simple steps. No code required.
Natural language compiles to a typed IR over the live tool registry; “open in code” for YAML + MCP.
Consume any external MCP server (scoped, revocable, SSRF-guarded) — and expose your agent as an MCP server to any MCP-compatible client.
Describe an agent in your own words. We draft it as plain steps from a hand-tested template and personalize it — you confirm before anything is created.
A constrained compiler — not open codegen — turns natural language into a typed, validated agent definition over the live TOOL_REGISTRY. A malformed agent is unconstructable.
name: Nudge Non-Repliers triggers: - kind: schedule cron: "0 9 * * 1" steps: - tool: get_tasks - tool: create_reminder guards: reminders: { level: auto_act_limited }
A named agent that introduces itself in your thread wearing a verified badge, does exactly the steps above, and asks before doing anything risky. Edit it any time in plain words.
Drafts from the closest curated template and personalizes the parameters. You confirm before anything is created.
Hand-authored, validated agents over the in-house tools — zero compiler risk. Every agent defaults to draft-only until you grant more.
Installed agents — each a verified principal with its own Ed25519 signing key. Inspect, test-run, or revoke. Real rows created by the API.
Scoped server-to-server keys for the REST API (/v1/*). The MCP surface uses OAuth-minted tokens, not direct keys. Shown once on create — store them securely. Rotate or revoke any time.
Let your agents use tools from other apps — and let any MCP-compatible client talk to your agents. Each connection is scoped and revocable.
Each agent is callable as its own OAuth-gated MCP server, and can reach external MCP servers through a scoped, SSRF-guarded, revocable broker.
Third-party apps and MCP-compatible clients you've granted access to via OAuth. Revoke any time.
Get a signed POST when events happen. SSRF-guarded; the HMAC signing secret is shown once.
Every signed POST Telbox sends — event, HTTP status, response time, retries.
Six tools. OAuth 2.1. Scoped + consented. Register via Dynamic Client Registration, complete the device flow, call the MCP surface with a scoped bearer. Clients on the trusted-software allowlist auto-approve.
📖 Open the full API reference & guides ↗| MCP tool | What it does | Scope |
|---|
Create → run → inspect → talk to the world. Every call is a real endpoint the console itself uses.
# set these once — create a key in the console (Settings → API keys)
export TELBOX=https://api.telbox.ai
export KEY=tb_live_…# 1 · install an agent from a template → returns its id
curl -X POST $TELBOX/v1/agents/from-template \
-H "Authorization: Bearer $KEY" -H 'Content-Type: application/json' \
-d '{"template_id":"vip-watcher"}'# 2 · or build one from a typed IR (what Studio ships)
curl -X POST $TELBOX/v1/agents \
-H "Authorization: Bearer $KEY" -H 'Content-Type: application/json' \
-d '{"name":"Nudge","triggers":[{"kind":"schedule","cron":"0 9 * * 1"}],
"steps":[{"id":"s1","type":"tool","tool":"get_tasks"},
{"id":"s2","type":"tool","tool":"create_reminder",
"args":{"title":{"prompt":"follow up with non-repliers"}}}],
"guards":{"capabilities":{"reminders":{"level":"auto_act_limited"}}}}'# 3 · test-run it (durable, side-effect aware) and read the trace
curl -X POST $TELBOX/v1/agents/$AGENT_ID/test-run \
-H "Authorization: Bearer $KEY" -H 'Content-Type: application/json' \
-d '{"prompt":"Give me a quick status of this thread."}'
curl $TELBOX/v1/agents/$AGENT_ID/runs -H "Authorization: Bearer $KEY"# 4 · call the MCP surface (global server, or a published agent's own /mcp)
curl -X POST $TELBOX/mcp \
-H "Authorization: Bearer $KEY" -H 'Content-Type: application/json' \
-d '{"method":"tools/call","params":{
"name":"send_message",
"arguments":{"thread_id":"THREAD_ID","content":"on it ✅"}}}'Every published agent is also its own OAuth-gated MCP server at POST /v1/agents/{id}/mcp — point any MCP-compatible client straight at it. Full machine-readable schema at /openapi.json.
Metered on real cost. Subscriptions buy limits + SLA, never entry. The paying customer is the business; consumers get agents inside their existing plan.
Developer-tier pricing is indicative and not final — calibrated to production COGS at ~75% gross margin. Marketplace: 85% to creators.
Drag triggers, tools, and logic onto the canvas and connect them into a working bot — no code. Every block is a real, verified capability; risky actions ask first.
A visual compiler over the live TOOL_REGISTRY. The canvas graph compiles to a typed AgentDefinitionIR — open it in code or ship it as a verified agent.
The same operations in your language — copy-paste against your key. Official SDK packages ship at GA; these call the REST + MCP surface directly today.