Two different kinds of “it should also know how to…”
When an agent cannot do something you need, it is almost always one of two problems, and they have different answers.
It cannot reach the system. Your ticketing tool, your metrics stack, your internal API. This is a tools problem — the agent needs a new capability.
It does not know your procedure. How your team does a release, what your incident checklist is, which conventions your repository follows. This is a knowledge problem — the agent has the capability but not your way of using it.
Forge answers them separately, because conflating them produces agents that either know too little or carry an enormous prompt.
Tools: your own MCP servers
MCP servers are attached per character, so the analyst can reach your metrics stack while the writer cannot. The definitions are validated when you save the agent — a bad URL or a duplicate name is rejected at the API boundary rather than surfacing as a confusing failure mid-run — and forwarded to the runtime when a turn is dispatched.
This is how an agent reaches something we have never heard of. You do not need us to build an integration first.
Knowledge: your own skills
A skill is a runbook — a procedure written down, that an agent can load when it needs it.
The design decision that matters here is what does not happen: skills are not inlined into every system prompt. Only a short index of ids and one-line descriptions goes in, and the agent fetches the full text when it decides a skill is relevant. Prompts stay cheap, agents stay focused, and a role never loads a runbook that has nothing to do with its job.
Organisation skills are additive: they sit alongside the built-in ones and cannot take over a built-in id. Adding your procedures cannot silently change behaviour that already worked.
What this is not
- Not a plugin marketplace. These are your tools and your runbooks, scoped to your organisation.
- Not a fork. Extending happens through configuration; you stay on the same platform everyone else is on.
- Not unbounded. Tools remain subject to the same egress and credential rules as everything else — an agent reaching your system still never holds the key.
Under the hood
- Character Agents — what a tool or skill gets attached to
- Secrets & Egress — how your systems get reached without handing over credentials
- Custom Roles — when the gap is the role itself rather than its tools