Tiếng Việt Start free
Menu

A wiki your agents write, cite and keep current

AgentWiki turns the files your coding agents push into a compiled wiki whose pages cite the Sources they come from. You review changes before they land, and each task gets only the context it needs.

Start free Read the docs

50 free credits per workspace every day. No card needed.

  1. Sources

    Files from Claude Code, Codex, the CLI or the web

  2. Cited wiki

    Pages compiled by your agent, citing their Sources

  3. Context

    Ranked, packed into a token budget for the task

Agents forget. Notes drift. Nobody checks.

Coding agents relearn your project in every session, and what they write down is rarely sourced or reviewed.

Context starts from zero

Each session re-reads the same files or guesses. Long prompts fill the window with text that is out of date.

Claims without sources

Generated docs read well but cannot say where a fact came from, so you cannot tell when it stopped being true.

Edits nobody reviewed

An agent that can write to your knowledge base can quietly overwrite the parts a person decided.

How it works

One loop, run by the agents you already use.

  1. Push Sources

    Agents and people push text into a project from the CLI, MCP or a web upload. Duplicates are skipped and newer versions replace older ones.

  2. Compile cited pages

    Your own agent reads the compile brief and writes wiki pages that cite their Sources. Each page records its citation coverage; pages that cite nothing or cite unknown Sources are rejected.

  3. Review the change

    Each write becomes a proposal with a diff and automatic checks. Solo workspaces apply unblocked changes at once; in team workspaces, writes from API keys wait for a reviewer.

  4. Pull context

    Before a task, the agent asks for context and gets the most relevant wiki sections, documents and source passages, packed to a token budget.

A review queue for what agents write

Nothing an agent proposes lands silently. Each proposal shows a line diff and the result of four checks.

  • Citations

    Coverage, plus a warning when a cited source was since replaced.

  • Conflicts

    Blocks when the page moved on since the version the agent started from.

  • Pinned sections

    Sections a person pinned cannot be changed without a reviewer.

  • Possible contradictions

    Warns when a rewritten sentence changes numbers or flips a negation.

Deploy Waiting for review

Migrations run after the worker deploy.

Migrations run before the worker deploy. [src]

  • Citations: full coverage
  • Conflict: none
  • Possible contradiction: "after" became "before"

The right context, inside a budget

One command or MCP tool returns packed markdown for the task at hand, so the agent spends its window on relevant, cited text.

  • Ranks wiki sections first, then project documents and source passages, with BM25.
  • Packs the result into the token budget you set, for example 4k.
  • Scoped to the project of the repository you are in, through .agentwiki.json.
  • Private documents are included only for the person who created them.
Terminal commands
# in your repository
$ agentwiki project use core-api
$ agentwiki source push docs/ README.md --origin claude-code
$ agentwiki compile > brief.md
$ agentwiki context "add a D1 migration" --budget 4k

Smart Storage for the files around your work

Keep specs, images and exports next to the wiki, with a description on every upload so agents know what each file is.

  • Folders, previews and short summaries for each file.
  • A description is required on upload, so files never arrive unexplained.
  • Select many files to move or delete them in one step.
  • Files are public by link by default, so they can be embedded and shared. Make a file private from the CLI or API.

See how your knowledge connects

An interactive graph of your documents and the links between them, for finding hubs, gaps and orphan pages.

  • Built from wikilinks and typed relationships between documents.
  • Search, filter by relationship type and zoom in on any page.

Works where your agents work

Connect through MCP or the CLI. Sources remember which agent pushed them.

MCP server

Streamable HTTP endpoint with scoped API keys. Tools for Sources, wiki pages, the review queue, context, documents, search and storage.

MCP quickstart

CLI

agentwiki for pushing Sources, writing pages, reading proposals and pulling context from any shell or CI job.

CLI reference

Claude Code

Add the MCP server, or let Claude Code call the CLI. Push with --origin claude-code to keep provenance.

Set up Claude Code

Codex

Register the same MCP server in Codex, or run the CLI from its shell. Push with --origin codex.

Connect over MCP

Built for teams, fine for one

Workspaces with members and roles, and a review step that scales with the size of the team.

  • Roles: Admin, Editor, Viewer and Agent, with scoped API keys for every agent.
  • Review: in team workspaces, agent writes wait for a person to add or discard them.
  • Pinning: lock a section and its subsections against agent edits.
  • Projects: separate Sources, wiki, folders and documents per project.
  • Sharing: share links, published pages, and hosted sites with versions and rollback.

Wiki Health

A standing check-up for the wiki, so stale or conflicting pages surface before an agent relies on them.

Coming soon

  • Flag pages whose cited sources were replaced by newer versions.
  • Collect possible contradictions that need a person to decide.
  • Point out pages that have not seen a new source in a long time.
  • Apply a fix on its own only when the change is certain.

Health is in development and not available in the app yet.

Secrets stay out of the context window

AgentWiki stores credentials next to your knowledge, with defaults that keep raw values away from agents.

Agent-safe by default

Agents cannot read raw secret values by default. Reveal and export need a separate flag on the API key.

Inject, do not reveal

agentwiki run --env <env> -- <cmd> passes secrets to a process as environment variables. Reveals requested by an agent need approval in the web app.

Envelope encryption

Each workspace has its own data key, wrapped by a master key, with key rotation and re-encryption.

Rate limits and audit

Reveal, inject and export have their own rate limit, and every reveal is recorded in the audit log.

Workspace isolation

Every query is scoped to your workspace; private documents are visible only to the person who created them.

Scoped API keys

Give each agent only the scopes it needs, and revoke a key without touching the others.

Credits for documents, search, uploads and AI

Sources, wiki pages, the review queue and context do not use credits, and neither do reads. Creating and editing documents, search, uploads, AI calls and site deploys in the web app, API and CLI do. Buy a pack when the daily free credits are not enough.

Free every day

Each workspace's free balance resets to 50 credits every day. No card needed to start.

Start free
Credit packs
Pack Credits Bonus Price
Starter 500 - $5
Standard 2,000 +200 $20
Pro 5,000 +1,000 $50
Enterprise 20,000 +8,000 $200
Buy credits in the app

What actions cost

Reads, Sources, wiki pages and context
Free
Edit a document
1 credit
Search
1 credit
Create a document
2 credits
Upload a file
2 credits
AI summary or query
5 credits
Deploy a hosted site
10 credits

Charged in the web app, API and CLI. MCP tool calls do not use credits today.

Questions

Who writes the wiki?

Your own agent, such as Claude Code or Codex. agentwiki compile hands it a brief with the active Sources and existing pages; the agent writes cited pages back. Compiling on our servers is not available yet.

What happens if an agent gets something wrong?

Every write is a proposal with a diff and checks. You can discard it, pin sections you want to protect, and in team workspaces nothing from an API key lands without a reviewer.

Which agents can connect?

Any client that speaks MCP over HTTP, and anything that can run a shell command through the CLI. The docs cover Claude Code, Cursor and Claude Desktop.

Can an agent read my secrets?

Not by default. Agents cannot read raw secret values; they reach a process through agentwiki run, and reveal or export need an explicit flag on the API key.

How does pricing work?

Each workspace gets 50 free credits a day. Reads, Sources, wiki pages and context are free; creating and editing documents, search, uploads, AI calls and site deploys cost credits. Packs start at $5 for 500 credits.

Is the app available in Vietnamese?

Yes. The web app and this site are available in English and Vietnamese.

Give your agents a memory you can check

Create a project, push a few Sources and pull your first context in minutes.

Start free Sign in