Casefile

One case file per task — shared by every session, agent and person that touches it.

Casefile keeps a case file for every task: the decisions, failed attempts, findings and open questions that agents write down as they work. When a session ends, the next one — the same agent, an agent from another vendor, or you — reads the latest summary and the open questions and carries on. It is a self-hosted MCP server and a web board, MIT-licensed; it assigns and schedules nothing.

For anyone whose agents work on tasks longer than one session. Free to use. Made for Claude Code; Codex, Cursor and any other MCP client connect the same way.

Install on macOS / Linux

curl -fsSL https://raw.githubusercontent.com/azimov777/casefile/main/install.sh | sh

Install on Windows (PowerShell)

irm https://raw.githubusercontent.com/azimov777/casefile/main/install.ps1 | iex

All you need is Docker. The board opens at http://localhost:8080, and the installer connects Claude Code and Codex by itself — a plugin and an OAuth sign-in, no token to copy. Casefile updates itself to each new release: it checks once an hour and whenever Docker starts. A small updater service does it and holds the Docker socket for that; you can turn it off.

Or let your agent do it. Paste this into Claude Code, Codex or Cursor:

Install Casefile for me by following https://raw.githubusercontent.com/azimov777/casefile/main/docs/agent-install.md

Why

A session ends or the context fills up, and the next agent starts from scratch: re-reading the code, re-trying what already failed, re-asking what you already answered.

Casefile gives every task a case file — an append-only log the agent writes as it works.

• Hand-offs that survive a fresh context. The next agent reads the latest summary, the open questions and an index of the case, then carries on. No re-discovery.

• Built for agents, over MCP. Agents create and split tasks, record decisions and dead ends, ask you questions, and close with a verdict on every check.

• You stay in the loop. A live board and task pages show what every agent is doing. Answer questions, leave remarks and hand each agent its own access — right from the browser.

• Guardrails, not bureaucracy. No closing without a summary and a passed verdict per check; no starting a blocked task. Nothing else — no sprints, no estimates, no automation.

• Yours, on your machine. Runs locally in Docker and listens on localhost only. Nothing leaves your computer — unless you turn on sign-in and put it on your own server for your team (Network mode).

Related MCP server: meridian

How it's different

• Not a notes file. A CLAUDE.md or handoff.md gets overwritten: the attempt that failed two days ago disappears, and two sessions edit the same file. A case file is append-only — a correction is a new entry that points at the old one. Keep CLAUDE.md for per-repo rules; Casefile is per task.

• Not a memory server. Memory MCPs recall facts by similarity. Casefile recalls nothing clever: it is a work log per task, read in a fixed order — card, latest summary, open questions, index, then only the entries you need.

• Not an issue tracker with MCP bolted on. An issue is a description and a thread anyone can edit. Case entries are typed and never edited, and the tracker refuses writes that would break the record.

• Not an orchestrator. It never starts agents, runs timers or moves tasks by itself. Handing out work and noticing a dead session stay with you and your agent harness.

Connect your agent

The installer connects Claude Code and Codex by itself. It installs the Casefile plugin — the skill and a connection to this installation's MCP address — and signs them in with OAuth: no token is copied and none lands in any file. If the installer had no terminal for the sign-in, it prints the commands; run them once:

claude mcp login plugin:casefile:casefile
codex mcp login casefile

The installer asks before it changes anything of another program. The plugin step edits files that are not Casefile's: Claude Code's ~/.claude/settings.json (one line, "autoUpdate": true, inside extraKnownMarketplaces.casefile), Codex's ~/.codex/config.toml, the manual MCP entries casefile and tracker that point at this installation, and ~/.agents/skills. It prints that list and asks [y/N] on your terminal (also under curl … | sh); N skips the whole plugin step, leaves those files alone and installs the service as usual. Before the first change to a file it saves a copy next to it — settings.json.casefile-bak, config.toml.casefile-bak, kept and never overwritten — and it writes only the line it needs, leaving the rest of the file as it was. The copy holds whatever the file held, a token of a manual entry the installer removes included: delete it when you no longer need it. These variables go in front of sh:

• CASEFILE_SKILL=0 skips the plugin step; CASEFILE_SKILL=1 answers the question in advance. Without a terminal (an agent, CI) nothing is asked and the step runs, as it does with 1.

• CASEFILE_PLUGIN_AUTOUPDATE=0 installs the Claude Code plugin without autoUpdate; update it yourself with claude plugin update casefile@casefile.

The same commands sign an agent in again after you disconnect it on the board's Access screen (in a Claude Code session, /mcp → casefile → Re-authenticate). On your own machine the sign-in needs no password: Claude Code acts as claude, Codex as codex.

The skill comes with the plugin. Connecting gives the agent the tools; the Casefile skill teaches it how to use them. The installer puts it into Claude Code, Codex, Hermes and other agents it finds on the machine, and prints one line per harness. An installation made before v0.8.0 has no skill and no plugin, and the hourly self-update does not add them: it updates only the service. Run the install line again, or install just the plugin and the skill without touching the service (add CASEFILE_URL=<address> for a server that is not on this machine):

curl -fsSL https://raw.githubusercontent.com/azimov777/casefile/main/install.sh | CASEFILE_SKILL_ONLY=1 sh

How to check whether an agent has the skill, and the commands for each harness, are in step 4 of the agent guide.

Claude Desktop gets an extension for its chat. The chat of the Claude desktop app (macOS, Windows) does not load the Claude Code plugin; it gets Casefile from casefile.mcpb, attached to every release. The installer downloads and opens it when it finds Claude Desktop; by hand, double-click the file. It signs in with OAuth, needs no Node.js and works while the installation runs (details).

Hermes signs in with OAuth (auth: oauth in ~/.hermes/config.yaml; the installer prints the block; not yet checked against a real Hermes). Clients without OAuth use the agent key. Scripts and any client that cannot sign in with MCP OAuth connect over streamable HTTP at the MCP address with the agent key in a header. The installer prints the key in their blocks; read it again with docker compose run --rm --no-deps -T agent-token cat .secrets/agent-token in ~/casefile. Such a client that takes an mcpServers JSON uses this — with the key and, if your installer printed a different address, that address instead:

{
  "mcpServers": {
    "casefile": {
      "type": "http",
      "url": "http://localhost:8100/mcp",
      "headers": {
        "Authorization": "Bearer <token>"
      }
    }
  }
}

Over stdio, as an alternative. Streamable HTTP above is the main way in. A client that can only launch a command and talk to it over stdin/stdout gets the same server that way: it starts a short-lived container of your installation, attached to the installation's database — same tools, same key, same case. There is no OAuth over stdio: the agent key goes in the client's environment, not on the command line:

claude mcp add --scope user casefile-stdio --env TRACKER_MCP_TOKEN=<token> -- \
  docker compose -f ~/casefile/docker-compose.prod.yml run --rm --no-deps -T \
  -e TRACKER_MCP_TOKEN mcp python -m app.mcp --stdio

The installation has to be up: the stdio process brings no database of its own. Each client session is a process of its own, so HTTP stays the lighter choice wherever the client supports it. An installation image older than the stdio mode answers unrecognized arguments: --stdio — update it first.

A second agent, without the terminal. Connect an agent on the board shows this installation's MCP address and ready-made snippets. Access lists every way in, in three sections: agent connections (OAuth sign-ins), agent keys and sign-in sessions — yours, or every one of the installation if you are an administrator: who it speaks for, who issued it and when it was last used. A connection is disconnected there and a key revoked. For a client without OAuth you register an agent, issue its own key there and copy the snippet with the key already in it — shown once. Give each agent its own sign-in or key and its case entries are signed with its name. On a shared installation every person does this for their own agents, without the administrator, and sees only the connections and keys they issued or that speak for them.

Tell it what to do

Connecting the agent is only the wiring. Once it is done, tell the agent these, word for word.

Have work to hand over? Say:

File tasks in Casefile for my work: a project for it if there is none yet, and tasks with all their sections and checks, each small enough for one agent to finish in one go, each naming its environment in context — where the work lives and how to check it is done. Don't start the work itself; if I haven't described it yet, ask me.

Then, in a new agent session, say:

Carry out the tasks for this work from the Casefile tracker. Hand them to agents, one task per agent, to save your own context, and give them cheaper models where those cope.

The two phrases go to different agent sessions: the second agent starts with a clean context and knows only what is in the task cases. The board carries the same phrases, with copy buttons, on its /start page.

Tools

Every MCP tool the server offers, grouped by area (app/mcp/tools/):

Tasks

• get_task — returns everything about one task in a single call: card, parent and children, links, computed features, latest summary, open questions, unresolved remarks, its discussions, case index and transition targets

• search_tasks — searches tasks by a query-language string, by separate conditions, or by both

• create_task — creates a task in backlog in an area of its project (area_required without one), optionally as a child of a parent task

• update_task — changes the given fields of a task; fields left out stay as they are

• transition — moves a task to another status along the fixed transition table

• close_task — closes a task: files entries, verdicts and the final summary and moves it to done, in one transaction

• move_task — moves a task, or each task of a list with an outcome per key, to another project with a reason; its previous key keeps leading to it

Case

• read_entries — returns entry bodies of one task's case, with payload, in number order

• add_summary — files a summary: the handover note of a case, in four parts

• add_entry — files an entry without payload: a decision, attempt, finding, artifact, remark or note

• ask — files a question to registry participants in a discussion; asked about a task, it opens the discussion and attaches the task in the same call

• answer — answers a question of a discussion or of a task's case, or closes one still unanswered as withdrawn or replaced by a later question, with a reason; an answered question cannot be withdrawn

• resolve — resolves a remark on a task: its outcome and where the work went

• add_verdict — files the outcome of one review check

• read_project_entries — returns entry bodies of one project's, area's or discussion's case, with payload, in number order; filters by number, type, status and a substring of the title or body

• add_project_entry — files a decision, finding, artifact or note in a project's or an area's case, or a note in a discussion's case

Discussions

• add_conclusion — files a discussion's conclusion: what is decided, superseded and still open

• close_discussion — closes a discussion with its final conclusion, in one transaction

• list_discussions — lists discussions with filters by project, status, turn and attached task: address, title, status, turn, open questions, closing time

Files

• request_file_upload — returns a one-time link and the curl -T command that uploads a new file to a project; the bytes go from disk to the tracker and never through the model

• request_file_replace — returns a one-time link and the curl -T command that replaces the content of a file under the same id

• read_file — returns a file: an image as image content the model sees, any other file as its card and a one-time curl -o link

• list_files — lists files as cards with search by name and filters by project and media type, sorted by time added or opened, size or type

• delete_file — erases a file's bytes and keeps its tombstone: name, type, size, who deleted it and when

• delete_files — erases every file a filter picks (project, name, type, unopened for images not opened for the term of days) once the confirmed count matches

Links

• link — links two tasks and files link_added in both cases, or attaches a task to a discussion

• unlink — removes a link and files link_removed in both cases, or detaches a task from a discussion

Projects & participants

• get_project — returns one project by its key, or an area by its address in the same shape: key, title, description, current attribute values, the decisions and findings in force and the index of the rest of its case

• list_projects — lists the installation's projects the agent may use: key, title and archive time; archived ones only when asked

• list_participants — lists the participant registry: the possible addressees of a question

• create_project — creates a project

• update_project — changes a project's title and description, recording each change in its case

• archive_project — archives a project with a reason, freezing it and its tasks against changes

• restore_project — restores an archived project with a reason

• set_attribute — sets the value of a project attribute, creating it or changing it with a reason; the history stays in the project's case

• remove_attribute — removes a project attribute with a reason, filing its last value in the project's case

• register_participant — registers a human or a permanent agent

• update_participant — changes a participant's description

Journal

• wait_journal — returns journal entries after a sequence number, waiting for new ones

Everyday

Ports and other settings live in ~/casefile/.env — see .env.example.

Updates

A new version of Casefile is a release: a git tag vX.Y.Z with its images on ghcr.io under the version and under the stable channel. Every installation follows stable by default. It checks when Docker starts and then once an hour, at a slightly random minute, so a release reaches it within about an hour and ten minutes, with nothing to restart. Commits to main without a tag never reach an installation. The update recreates the Casefile containers and keeps your data in its volumes. An agent in the middle of an MCP call when that happens gets a dropped connection and has to retry.

The updater holds the Docker socket. To replace containers the updater service mounts /var/run/docker.sock, which is root access to the machine; Casefile gives it to this one service on purpose (docker-compose.prod.yml says why). The installer tells you this before it installs and again at the end. To turn auto-update off, put CASEFILE_AUTO_UPDATE=false into ~/casefile/.env (exactly false: any other value leaves it on) and run docker compose up -d --no-deps updater in ~/casefile. The service then idles, so the socket stays mounted; docker compose stop updater stops it until the next docker compose up or run of the installer. With auto-update off the board tells you when a new release is out: a note at the bottom of the side panel links to the release notes, and you update by running the install line again. That note comes from a read of the latest release on GitHub, at most once an hour; TRACKER_RELEASE_CHECK=false in .env turns the read off, and the note with it.

The first release on this channel was 0.2.0. An installation from before it (on latest) moves to stable by itself the next time Docker starts, and from then on checks every hour. A release can also bring a new updater: the update to that release is still done by the old one, which is then replaced by itself, so what a new updater adds (such as the rollback below, new in 0.3.0) covers updates from the next release on. If you set CASEFILE_VERSION=latest in .env yourself, remove the line to follow releases.

CASEFILE_UPDATE_INTERVAL sets how often to check (hours, or 30m; 0 means only when Docker starts). If a release fails to start, the installation goes back to the version it ran before and does not try that release again; the next release is installed as usual. docker compose logs updater tells what happened.

A release that changes the database schema costs one more step. Before installing it, the updater takes a snapshot of the database (pg_dump -Fc, the same format as backup and restore). The snapshot goes into a Docker volume of the installation, casefile_updater-snapshot, as before-update.dump, and takes about as much space as a manual backup. If that release then fails to start after changing the schema, the database goes back to the snapshot before the previous version starts again. So the schema and the data are exactly as they were before the update, and a manual docker compose up -d works as usual. The price: anything written between the snapshot and the rollback is lost. That window is the failed start, up to a few minutes. While the new version is being brought up, the old one keeps answering for a few seconds. The snapshot is not deleted afterwards, neither after a good update nor after a rollback: it stays in the volume as a copy of the database from before the update, and docker compose logs updater says where. There is one copy: the next update that changes the schema replaces it, and a snapshot that fails never replaces the earlier one. How to restore from it: backup and restore. If the snapshot cannot be taken, that release is not installed this time. If it cannot be restored, the previous version runs on the new schema, and the log says how to copy the snapshot out.

Network mode

Out of the box Casefile listens on localhost only, and you type nothing: the installation creates an administrator account for you (owner@localhost) and the board signs into it by itself. To reach it from other machines, turn on sign-in: then everyone signs in with their own email and password, and every entry is signed by the person who made it. It is one team per installation — everyone signed in sees every task. The only role is the administrator flag, and all it opens is managing people.

1. Add to ~/casefile/.env:

   CASEFILE_LOGIN=password            # everyone signs in with email and password
   CASEFILE_BIND=0.0.0.0              # publish the board and MCP beyond localhost
   TRACKER_MCP_PUBLIC_URL=http://<server>:8100/mcp   # what agents on other machines use

2. Run docker compose up -d in ~/casefile.

3. Give yourself a password. The administrator account the installation made has none; this prints a generated one, once:

   docker compose run --rm api python -m app.cli account-password --email owner@localhost

   Add --set-password to type your own instead (12 characters at least, asked twice, no echo). Your email can be changed too: account-update --email owner@localhost --new-email you@example.com.

4. Add your teammates — on the board, or on the server:

   docker compose run --rm api python -m app.cli account-create --email alice@example.com --name alice

   It prints Alice's password once; hand it to her. --admin makes her an administrator too. account-list shows everyone, account-update --disable locks a person out and revokes their sessions and every connection and key they issued to their agents (their past entries stay signed with their name), and account-password resets a forgotten password. Casefile sends no mail: there is no address confirmation and no reset link.

The board at http://<server>:8080 now opens with a sign-in screen. Agents sign in with OAuth behind TLS (below); Connect an agent shows the address from TRACKER_MCP_PUBLIC_URL. Each person's agent machine gets the plugin, the skill and the sign-in with one line — no Docker and no service installer there:

curl -fsSL https://raw.githubusercontent.com/azimov777/casefile/main/install.sh | CASEFILE_SKILL_ONLY=1 CASEFILE_URL=https://mcp.casefile.example.com/mcp sh

The steps for the agent are in the agent guide.

Coming from the owner password. An installation locked with TRACKER_PASSWORD_HASH before accounts existed keeps working after the update: sign-in turns on by itself, and the old password becomes the password of the administrator account owner@localhost — sign in with that email and the same password. The hash in .env is no lock any more; it is only carried over once, and a password you set later is never overwritten by it.

Plain HTTP is a hole. Without TLS the passwords, the session cookies, the keys and the agents' tokens cross the network in clear text for anyone on the path to read. Casefile does not do TLS itself. Anywhere beyond a network you trust, keep CASEFILE_BIND=127.0.0.1 and put a reverse proxy with TLS in front of both ports — for example Caddy, which gets the certificates and sends X-Forwarded-Proto by itself:

casefile.example.com {
    reverse_proxy 127.0.0.1:8080
}
mcp.casefile.example.com {
    reverse_proxy 127.0.0.1:8100
}

with TRACKER_MCP_PUBLIC_URL=https://mcp.casefile.example.com/mcp. A proxy that sends X-Forwarded-Proto: https gets the session cookie marked Secure.

Agents sign in with OAuth. With sign-in on and TRACKER_MCP_PUBLIC_URL on https, a client that speaks MCP OAuth (Claude Code, Codex) needs only the address, no token. It opens a page on the MCP host: the person signs in with email and password and picks which agent the client acts as. By default it is their own agent, created on the first sign-in: Alice's Claude Code signs its case entries as claude_alice, her Codex as codex_alice (agent_<name> for other clients). The list holds their own agents and agents without an owner, never someone else's. The connection counts as issued by that person: they see and disconnect it on Access, an administrator sees all. Disabling the person's account revokes every connection and key they issued — their agents' too: a connection cannot renew itself and has to be signed in anew once the account is back. Harnesses without OAuth, and a journal watcher running between sessions, use a key the person issues to their agent on Access. What the proxy must do:

• send the whole MCP host to port 8100, not only /mcp: sign-in also uses /.well-known/…, /register, /authorize, /token and /oauth/… at the root of the host named in TRACKER_MCP_PUBLIC_URL;

• keep the client's Host header. Caddy does; nginx needs proxy_set_header Host $host;. With sign-in on, MCP answers only on the host of TRACKER_MCP_PUBLIC_URL, and a request under another name gets 421.

Over plain http beyond localhost there is no OAuth at all — agents connect with a key from Access. On a machine of your own (sign-in off, ports on localhost) an agent is signed in at once, without a page: Claude Code as claude, Codex as codex, any other client as agent, each created when first needed.

Name your proxy. Sign-in tells guessers apart by address (see Guessing below). Behind a proxy every request arrives from the proxy, so the board must be told which address is the proxy; only then does it take the browser's address from the proxy's X-Forwarded-For. From anyone else that header is ignored — anybody can write it. Add to ~/casefile/.env:

CASEFILE_TRUSTED_PROXIES=172.18.0.1   # the address the board sees the proxy come from

and run docker compose up -d. That address is not the proxy's own: it is whatever Docker shows the board. With the proxy on the same machine and CASEFILE_BIND=127.0.0.1, it is the gateway of the installation's network, on Linux and Docker Desktop alike:

docker network inspect casefile_default -f '{{range .IPAM.Config}}{{.Gateway}}{{end}}'

To check, run docker compose logs ui: each request line starts with the address it came from — with the proxy named, the browser's; without, the proxy's. Several proxies or a network go comma-separated (172.18.0.1,10.0.0.0/8). The network gets its address when it is created, so after docker compose down check the gateway again. Keep CASEFILE_BIND=127.0.0.1 behind a proxy: Docker Desktop shows every connection to a port published to the network as 192.168.65.1, so naming that address would let anyone claim any address. Without this line nothing breaks, but everyone behind the proxy shares one address — and one guesser holds everybody at "try again later" again.

What else to know:

• No sign-in, no network. With CASEFILE_BIND beyond localhost and sign-in off, the board refuses to start instead of handing the administrator key to the whole network; docker compose logs ui says why.

• Sessions. A sign-in lasts 7 days (TRACKER_SESSION_HOURS). A session is a token with a deadline, kept in the database: restarting the installation does not end it. Sign out revokes it at once, a changed or reset password ends the person's other sessions, and disabling an account revokes all its sessions and every connection and key the person issued to their agents: enabling it again brings none of them back.

• Guessing. Wrong passwords are counted per address and per email, within a minute. After 5 from one address, sign-in answers "try again later" to that address — the right password included — until the minute has passed; after 5 for one email, from wherever they come, that email waits the same way. Other people sign in as usual. On top of that the whole installation takes at most 100 wrong passwords a minute, to keep guessing from burning the processor; guessing spread over many addresses and many emails that reaches it keeps everyone at "try again later" for as long as it goes on (an IPv6 client counts as its whole /64). Open sessions and the agents, which use tokens, are not affected. A client is the address that opened the connection, or the one your named proxy reports (above). Docker Desktop hides the address of every connection from the network behind one of its own, so there the board tells clients apart only behind a proxy on the same machine.

Under the hood

Python 3.14 · FastAPI · PostgreSQL · MCP over streamable HTTP · React 19 · Vite · Tailwind. The backend sits at the repository root, the web UI in ui/. The web UI speaks English and Russian and the agent-facing texts are in Russian for now; design decisions live in the maintainer's tracker, so ask in an issue.

Contributing

Issues and pull requests are welcome — start with CONTRIBUTING.md. Commits need a sign-off (git commit -s): it certifies you have the right to submit the code, and CI checks it.

License

MIT

Hiring? I built Casefile and would be glad to hear about roles at Anthropic or OpenAI — reach me through GitHub.