Six capabilities that make the agent yours.
Tangison Agent is a self-hosted AI workforce that runs entirely on your hardware. It watches the channels you point it at, plans a sequence of skill calls to satisfy each request, and executes those skills inside a workspace you control. The model is yours, the prompts are yours, and the audit trail is yours. This page breaks down the six capabilities that make that possible, and shows how each one maps to a concrete operational benefit you can take to a compliance review or a budget meeting. Each feature was built to answer a specific failure mode of cloud-hosted AI: data leakage, vendor lock-in, hidden markups, opaque reasoning, and single-channel delivery. Tangison Agent is a project of Tangison, the independent brand behind tangison.com, and ships under an MIT license at github.com/tangison/agent.
Feature deep-dives
Each feature, one row, one business outcome
Six alternating rows, six honest answers to the question of why this capability matters in production. Read them in order, or jump to the comparison table below to see Tangison Agent lined up against a generic cloud AI offering. The feature set was decided in the open at tangison.com before a single line of agent code was written.
01 / 06
Sovereign by default
The agent process, the workspace, and the audit log all live on your hardware. No third party ever sees your prompts, your files, or your tool outputs. You own the server, the network, and the boundary. When a compliance officer asks where the data lives, the answer is one folder on one machine you control, with logs you can hand them on a USB stick.
Your prompts and tool outputs never leave a machine you own.
02 / 06
Bring your own model key
Plug in your OpenRouter API key. Pick the model that fits the job, from Claude to GPT to Mistral to local Llama. Tangison never resells credits, never marks up your usage, and never picks the model for you. You see the raw model name and the raw token cost on every task, with no middleman fee layered on top and no surprise bill at the end of the month.
Pay the model provider directly. No markup, no resale, no surprise bill.
03 / 06
Composable skills
Seven built-in skills ship in the box: web_search, vercel_deploy, send_message, read_write_files, run_code, fetch_url, and mcp_connector. Each one is a typed TypeScript module with a clear input and output contract. Add your own in TypeScript with a typed Skill<I, O> interface and the dispatcher picks it up on the next boot, no rebuild required and no remote code loading.
New behavior in an afternoon, with types and tests you can trust.
04 / 06
Multi-channel delivery
Ship the agent to Telegram, WhatsApp, and the web dashboard in one config block. Same brain, every channel, no extra bill. Each channel is a thin transport adapter that turns an inbound message into the same agent event. A conversation started on WhatsApp can finish on Telegram without losing context, audit log, or skill state, because the channel is just a wire.
One agent, every inbox your customers already use.
05 / 06
Open source, MIT licensed
Read every line. Fork it. Run it offline. Contribute back. No black boxes, no telemetry you did not write yourself. The repo lives at github.com/tangison/agent under an MIT license, with public issue tracking and a published changelog for every release. If you find a bug, you can fix it on a branch and open a pull request the same afternoon, with no enterprise sales call in between.
No vendor lock-in. The day you outgrow us, you keep the code.
06 / 06
Full operational control
Workspace isolation, per-skill network policies, and a complete task log. You see every input, every output, every side effect, and the audit trail is yours to keep, export, and replay. Per-skill network policies let you grant web_search outbound access while keeping read_write_files locked to a sandboxed folder, so a runaway skill cannot exfiltrate data or phone home to a host you did not approve.
Prove what the agent did, what it touched, and what it sent.
Honest comparison
Tangison Agent vs generic cloud AI
No puffery. Each cell is a short honest phrase about how the two approaches handle the same capability. Cloud AI is a good choice for some teams, and we are not going to pretend otherwise, but the trade-offs change when your data, your model spend, and your audit trail live on someone else's hard drive. Tangison Agent was built at tangison.com to make those trade-offs explicit, not to hide them behind a slick pricing page.
| Capability | Tangison Agent | Generic cloud AI |
|---|---|---|
| Data residency | Lives on your hardware, in a folder you control. | Stored in vendor cloud, region chosen by the provider. |
| Model choice | Any model on OpenRouter, plus local Llama. Per task. | A short list the vendor curates and prices. |
| Pricing model | You pay the model provider directly. Tangison charges for support, not tokens. | Per-token markup, seat fees, and usage caps set by the vendor. |
| Audit trail | Every input, output, and side effect logged to a file you own. | Limited to last 30 days in a vendor dashboard you cannot export. |
| Source code | MIT licensed, on GitHub. Read it, fork it, patch it. | Closed. You trust the vendor or you do not. |
| Channel multiplexing | Telegram, WhatsApp, and web share one brain, no extra bill. | Each channel is a separate integration with its own pricing. |
| Skill authoring | TypeScript Skill<I, O> interface. Add your own in an afternoon. | Plugin SDK in the vendor's language, with review queue and approval. |
| Offline operation | Runs fully offline with a local Llama model. No phone-home. | Requires constant connectivity to the vendor's API. |
Pick the features that matter to you
Tell Tangi which capability sold you, which one you want to test first, and what infrastructure you run. The waitlist opens directly to him on WhatsApp, and he replies within one business day with a self-hosting plan and a model recommendation.