Agent integrations

kontext is an MCP server: kontext mcp speaks JSON-RPC 2.0 over stdio. It answers both the stateless 2026-07-28 revision (server/discover, no initialize) and the legacy handshake for protocol versions 2024-11-05 to 2025-11-25. Every MCP client gets the same sixteen tools, prompts and resources.

What agents get

Tools (see the MCP reference for parameters):

Tool	When an agent should call it
ctx_brief	at the start of every task, with focus = the paths or topic it will work on
ctx_search	to find decisions, docs, history or memory about something
ctx_read	to open a hit at L0/L1/L2
ctx_why	before changing code whose reason is unclear (path, path:line, commit, symbol)
ctx_log	to see the decision timeline
ctx_drift	to see which entries a change left behind
ctx_threads	to read past agent sessions (Claude Code, Codex, Superset)
ctx_capture	when something durable was decided or learned
ctx_challenge	when an entry looks wrong: a challenge for a person to resolve
ctx_confirm	after checking that an entry still holds (recorded as an agent stamp)
ctx_verify	to run an entry's pinned verify command; an unpinned one is returned for the agent to run itself
ctx_inbox	to list, show, promote or drop captured candidates
ctx_prepare_commit	before every commit
ctx_init, ctx_init_submit	to bootstrap and deepen the repository's knowledge
ctx_api	to call an application API operation

Read tools (ctx_brief, ctx_search, ctx_read, ctx_why, ctx_log, ctx_drift, ctx_threads) are readOnlyHint: true. ctx_verify is destructiveHint: true and openWorldHint: true; ctx_api is openWorldHint: true. Tools re-published from MCP adapters (expose) keep their upstream annotations and appear next to these, as do tools declared in adapter config.

Structured results: ctx_brief, ctx_search, ctx_read, ctx_why, ctx_inbox and ctx_drift return structuredContent beside the text, and hits carry resource_link blocks to kontext://entry/{id}.

Prompts: kontext-init, kontext-commit, kontext-reflect (slash commands in Claude Code).

Resources: kontext://brief and kontext://entry/<id> for clients that browse resources.

Instructions: on connect the server tells the agent how to use the tools above. The first 512 characters stand alone (what kontext holds, call ctx_brief first, when to search) and the whole stays under 4 096 characters, the limits Codex and Claude Code read. ctx_brief carries alwaysLoad, so hosts that defer tools behind search still load it.

Hooks: kontext hook --host claude|codex|cursor delivers the brief at session start and the governing entries before an edit, opt-in with kontext connect <host> --hooks. See Hooks.

Behaviour of the server

• Requests are handled concurrently. A panic in one request answers -32603; the server keeps running.
• The repository is discovered from the working directory (or -C <dir>, or KONTEXT_DIR). The server also starts outside a repository, says so, and still serves calls with dir.
• Configuration edits (a new .ai/kontext.toml, a changed adapter) are picked up without a restart; a config that does not parse comes back as a warning in the next result.
• An index refreshed by another process (a CLI capture, kontext status, a second session) is reloaded before the next answer. The branch is read per request.
• notifications/cancelled stops the request and kills a running verify command.
• Adapter MCP servers are started on first use and kept for the lifetime of the session.

Clients

Client	Guide
Claude Code	Claude Code
Codex	Codex
Cursor, OpenCode, Superset, other MCP clients	Other clients

Without MCP

Everything is also a CLI command (kontext brief, kontext search, …), and kontext call <tool> '<json>' runs any tool locally, which is handy for harnesses without MCP support and for scripts.